diff --git "a/results.json" "b/results.json"
--- "a/results.json"
+++ "b/results.json"
@@ -48,6 +48,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"bn",
@@ -90,6 +97,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"de",
@@ -132,6 +146,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"en",
@@ -174,6 +195,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"es",
@@ -216,6 +244,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"fr",
@@ -342,6 +377,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ja",
@@ -531,6 +573,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ru",
@@ -573,6 +622,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"sw",
@@ -615,6 +671,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"te",
@@ -734,6 +797,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"zh",
@@ -770,12016 +840,10084 @@
     "score":0.3243323503
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1806181405
+    "score":0.2660357996
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4614154717
+    "score":0.5489237487
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3184146682
+    "score":0.3431691167
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4885651648
+    "score":0.5659025601
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.8
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1707650511
+    "score":0.2999810338
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4191623205
+    "score":0.5315905896
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1747239771
+    "score":0.3568113924
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4138755715
+    "score":0.5379558638
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.214232349
+    "score":0.3410327538
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4760523398
+    "score":0.5702449417
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3694909412
+    "score":0.4621824412
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6379394313
+    "score":0.6644677733
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3898301226
+    "score":0.5530727537
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6012929861
+    "score":0.7075859327
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.488423591
+    "score":0.6368037655
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7529549619
+    "score":0.8170495194
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2337674235
+    "score":0.3558328881
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4951544114
+    "score":0.5816613686
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.341066665
+    "score":0.3018683161
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6161709719
+    "score":0.5862734644
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.199697024
+    "score":0.3482763208
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4970728001
+    "score":0.5982327963
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.548967737
+    "score":0.452269216
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7095016995
+    "score":0.6590312745
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.367723046
+    "score":0.3747394322
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5881358889
+    "score":0.6079633657
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3278391831
+    "score":0.4420465241
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5487381665
+    "score":0.6372857982
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2012670144
+    "score":0.3120299438
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4903423287
+    "score":0.5421647326
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2590870164
+    "score":0.3139800299
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5931106206
+    "score":0.6305869448
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1789581673
+    "score":0.2121435327
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4444827041
+    "score":0.5165098934
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.175930196
+    "score":0.3965805608
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.329462033
+    "score":0.5197322727
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"mr",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"mr",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1834187267
+    "score":0.2594796679
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"mr",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.470878386
+    "score":0.4780086047
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"mr",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1744972384
+    "score":0.3009966401
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"mr",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4213302564
+    "score":0.5761529867
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"pa",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"pa",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3488186237
+    "score":0.2536168105
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"pa",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5904301709
+    "score":0.5193737639
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"pa",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3621810085
+    "score":0.2743904396
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"pa",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.51265991
+    "score":0.4918283752
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"pt",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"pt",
-    "task":"mmlu",
-    "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"pt",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2014349032
+    "score":0.4307107958
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pa",
+    "task":"translation_from",
+    "metric":"chrf",
+    "score":0.6547934264
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pa",
+    "task":"translation_to",
+    "metric":"bleu",
+    "score":0.3789232965
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pa",
+    "task":"translation_to",
+    "metric":"chrf",
+    "score":0.5490387026
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pt",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pt",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pt",
+    "task":"translation_from",
+    "metric":"bleu",
+    "score":0.3491027097
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.458360538
+    "score":0.5771246685
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4537177243
+    "score":0.3737440632
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6780859486
+    "score":0.6547073078
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2192993586
+    "score":0.2584319121
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5037493911
+    "score":0.5177022275
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3365435366
+    "score":0.371218137
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5803139095
+    "score":0.5948906549
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1629528612
+    "score":0.2583542451
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3934230515
+    "score":0.5157482031
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2569355275
+    "score":0.3458513734
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5756052088
+    "score":0.634706105
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.6
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3030045526
+    "score":0.3887890595
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5524023581
+    "score":0.6100201392
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.197839972
+    "score":0.3146346509
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4780950203
+    "score":0.5476489231
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1997540464
+    "score":0.2324983634
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4750614684
+    "score":0.497221173
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1234051854
+    "score":0.2976387517
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2799574404
+    "score":0.4761547661
   },
   {
-    "model":"anthropic\/claude-3-haiku",
-    "bcp_47":"zh",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"vi",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"vi",
+    "task":"translation_from",
+    "metric":"bleu",
+    "score":0.3027505857
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"vi",
+    "task":"translation_from",
+    "metric":"chrf",
+    "score":0.5506378818
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"vi",
+    "task":"translation_to",
+    "metric":"bleu",
+    "score":0.3708866541
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"vi",
+    "task":"translation_to",
+    "metric":"chrf",
+    "score":0.5846851624
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"zh",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1922766267
+    "score":0.2377253988
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4644593408
+    "score":0.5405656214
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2227195824
+    "score":0.2660263348
   },
   {
-    "model":"anthropic\/claude-3-haiku",
+    "model":"deepseek\/deepseek-chat",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2687577774
+    "score":0.3459946232
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2126059737
+    "score":0.2993048546
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4462765689
+    "score":0.5651988199
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1836371582
+    "score":0.3853695566
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3778102006
+    "score":0.602672086
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.4
+    "score":0.7
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0953670352
+    "score":0.2798537803
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3310320034
+    "score":0.5394779979
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.024360011
+    "score":0.3719179468
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2108567474
+    "score":0.539682577
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.8
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2306746115
+    "score":0.3207762021
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4506487927
+    "score":0.5555389401
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2707214129
+    "score":0.4227752207
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4862384022
+    "score":0.6601015066
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2371132106
+    "score":0.5548034204
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4776904622
+    "score":0.7071106777
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2328322481
+    "score":0.6432470265
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5236467405
+    "score":0.8278285651
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2281708898
+    "score":0.3223436235
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4658048059
+    "score":0.5379000389
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.302206669
+    "score":0.3061941236
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5004953235
+    "score":0.5918382188
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.5
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2485056339
+    "score":0.3297556296
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4746922491
+    "score":0.5925809306
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3589308099
+    "score":0.4619761505
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5387167593
+    "score":0.6734036273
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":1.0
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.176310372
+    "score":0.3412169553
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3718450169
+    "score":0.598919602
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1638320264
+    "score":0.3937102354
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4001947091
+    "score":0.6229439454
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2256160181
+    "score":0.3020610187
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4634218276
+    "score":0.5462026627
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2436068148
+    "score":0.299038365
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5064853457
+    "score":0.625735911
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.7
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2136903951
+    "score":0.1875636541
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4243463352
+    "score":0.5136106256
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0970742193
+    "score":0.3333377273
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2293867344
+    "score":0.4709407515
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"jv",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"jv",
+    "task":"translation_from",
+    "metric":"bleu",
+    "score":0.2789250445
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"jv",
+    "task":"translation_from",
+    "metric":"chrf",
+    "score":0.52614288
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"jv",
+    "task":"translation_to",
+    "metric":"bleu",
+    "score":0.2580648249
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"jv",
+    "task":"translation_to",
+    "metric":"chrf",
+    "score":0.574708573
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0799150208
+    "score":0.2487969868
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2893419537
+    "score":0.518708582
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0581409044
+    "score":0.2238578938
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2268165833
+    "score":0.4748109447
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.1
+    "score":1.0
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1097931752
+    "score":0.3823229705
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2693246745
+    "score":0.6453681393
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.056288688
+    "score":0.3548387061
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.1832547974
+    "score":0.5492435889
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2273037816
+    "score":0.3091555064
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4505835688
+    "score":0.5528775735
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3286476013
+    "score":0.3860807525
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5091531561
+    "score":0.6710753294
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.7
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1855609032
+    "score":0.2464304597
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4153884364
+    "score":0.5343201712
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.277880372
+    "score":0.3592456339
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4652042594
+    "score":0.5816925415
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.7
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0369694992
+    "score":0.2438930348
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.240137357
+    "score":0.5213646779
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0127960718
+    "score":0.2864351463
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2182424065
+    "score":0.6249321785
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.4
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0940291711
+    "score":0.3464968589
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2872804994
+    "score":0.5970733128
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0398689493
+    "score":0.2948086539
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2075817756
+    "score":0.5417485172
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.9
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.117904807
+    "score":0.2290327476
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3035503387
+    "score":0.5229250115
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0043914583
+    "score":0.3122409611
   },
   {
-    "model":"cohere\/command-r",
+    "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.1973427325
+    "score":0.4921734247
   },
   {
-    "model":"cohere\/command-r",
-    "bcp_47":"zh",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"cohere\/command-r",
-    "bcp_47":"zh",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.8
   },
   {
-    "model":"cohere\/command-r",
-    "bcp_47":"zh",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1818080543
+    "score":0.2150236607
   },
   {
-    "model":"cohere\/command-r",
-    "bcp_47":"zh",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4402625664
+    "score":0.4970978512
   },
   {
-    "model":"cohere\/command-r",
-    "bcp_47":"zh",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1788431133
+    "score":0.3603191861
   },
   {
-    "model":"cohere\/command-r",
-    "bcp_47":"zh",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2242902764
+    "score":0.5938509481
   },
   {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ar",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ar",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ar",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2660357996
+    "score":0.2350235637
   },
   {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ar",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5489237487
+    "score":0.528086246
   },
   {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ar",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3431691167
+    "score":0.2524231151
   },
   {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ar",
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5659025601
+    "score":0.3456705882
   },
   {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"bn",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.1
   },
   {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"bn",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"bn",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2999810338
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"bn",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5315905896
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"bn",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3568113924
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"bn",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5379558638
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"de",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"de",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"de",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3410327538
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"de",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5702449417
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"de",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4621824412
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"de",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6644677733
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"en",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"en",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"en",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.5530727537
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"en",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.7075859327
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"en",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.6422296046
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"en",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.81811999
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"es",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"es",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"es",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3558328881
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"es",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5816613686
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"es",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3018683161
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"es",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5862734644
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"fr",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"fr",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"fr",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3482763208
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"fr",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5982327963
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"fr",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.452269216
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"fr",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6590312745
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"hi",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"hi",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"hi",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3747394322
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"hi",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.6079633657
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"hi",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4420465241
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"hi",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6372857982
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"id",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"id",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"id",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3120299438
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"id",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5421647326
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"id",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3139800299
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"id",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6305869448
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ja",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ja",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ja",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2121435327
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ja",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5165098934
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ja",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3965805608
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ja",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5197322727
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"jv",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"jv",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2594796679
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"jv",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4780086047
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"jv",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3009966401
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"jv",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5761529867
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"mr",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"mr",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2536168105
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"mr",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5193737639
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"mr",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2743904396
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"mr",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4918283752
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pa",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pa",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.4307107958
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pa",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.6547934264
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pa",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3789232965
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pa",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5490387026
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pt",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pt",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pt",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3491027097
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pt",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5771246685
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pt",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3737440632
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"pt",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6547073078
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ru",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ru",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ru",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2584319121
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ru",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5177022275
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ru",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.371218137
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ru",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5948906549
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"sw",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"sw",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"sw",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2583542451
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"sw",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5157482031
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"sw",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3458513734
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"sw",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.634706105
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"te",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"te",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.6
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"te",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3887890595
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"te",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.6100201392
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"te",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3146346509
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"te",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5476489231
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ur",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ur",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2324983634
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ur",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.497221173
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ur",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2976387517
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"ur",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4761547661
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"vi",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"vi",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"vi",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3027505857
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"vi",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5506378818
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"vi",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3708866541
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"vi",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5846851624
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"zh",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"zh",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"zh",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2377253988
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"zh",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5405656214
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"zh",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2660263348
-  },
-  {
-    "model":"deepseek\/deepseek-chat",
-    "bcp_47":"zh",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3459946232
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ar",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ar",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ar",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2993048546
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ar",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5651988199
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ar",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3853695566
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ar",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.602672086
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"bn",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"bn",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"bn",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2798537803
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"bn",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5394779979
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"bn",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3719179468
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"bn",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.539682577
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"de",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"de",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"de",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3207762021
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"de",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5555389401
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"de",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4227752207
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"de",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6601015066
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"en",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"en",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"en",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.569880154
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"en",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.7141874172
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"en",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.6432470265
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"en",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.8278285651
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"es",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"es",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"es",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3223436235
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"es",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5379000389
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"es",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3061941236
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"es",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5918382188
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"fr",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"fr",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.5
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"fr",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3297556296
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"fr",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5925809306
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"fr",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4619761505
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"fr",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6734036273
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"hi",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"hi",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"hi",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3670345278
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"hi",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.6162795112
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"hi",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3937102354
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"hi",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6229439454
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"id",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"id",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"id",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3020610187
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"id",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5462026627
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"id",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.299038365
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"id",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.625735911
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ja",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ja",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ja",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.1875636541
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ja",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5136106256
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ja",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3333377273
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ja",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4709407515
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"jv",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"jv",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2789250445
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"jv",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.52614288
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"jv",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2580648249
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"jv",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.574708573
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"mr",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"mr",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2487969868
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"mr",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.518708582
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"mr",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2238578938
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"mr",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4748109447
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pa",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pa",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.4358296524
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pa",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.6732700357
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pa",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3548387061
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pa",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5492435889
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pt",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pt",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pt",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3091555064
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pt",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5528775735
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pt",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3860807525
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"pt",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6710753294
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ru",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ru",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ru",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2464304597
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ru",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5343201712
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ru",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3592456339
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ru",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5816925415
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"sw",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"sw",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"sw",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2438930348
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"sw",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5213646779
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"sw",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2864351463
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"sw",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6249321785
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"te",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"te",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.4
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"te",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3465514125
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"te",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.588783736
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"te",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2948086539
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"te",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5417485172
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ur",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ur",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2290327476
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ur",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5229250115
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ur",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3122409611
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"ur",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4921734247
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"vi",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"vi",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"vi",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2150236607
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"vi",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4970978512
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"vi",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3603191861
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"vi",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5938509481
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"zh",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"zh",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"zh",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2350235637
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"zh",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.528086246
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"zh",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2524231151
-  },
-  {
-    "model":"deepseek\/deepseek-chat-v3-0324",
-    "bcp_47":"zh",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3456705882
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ar",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ar",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ar",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2238237549
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ar",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4534366926
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ar",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2689055687
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ar",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4149416248
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"bn",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.2
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"bn",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"bn",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2443635406
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"bn",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4676159664
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"bn",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2223182846
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"bn",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3685961254
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"de",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"de",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"de",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2633055293
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"de",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.451378667
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"de",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2576565152
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"de",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3874723625
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"en",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"en",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"en",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3516668412
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"en",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4533116867
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"en",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.5868227988
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"en",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.7574063883
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"es",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"es",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"es",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2163623393
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"es",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4705179867
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"es",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2458237388
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"es",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.426538099
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"fr",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"fr",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"fr",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3182982487
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"fr",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5389072956
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"fr",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.47314841
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"fr",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6884710951
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"hi",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.2
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"hi",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"hi",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2751556219
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"hi",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4216960004
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"hi",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2827365983
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"hi",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5128198247
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"id",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.2
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"id",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"id",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.223787985
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"id",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.485186041
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"id",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3571004344
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"id",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6398491182
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ja",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.2
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ja",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ja",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.1711796281
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ja",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4108641598
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ja",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2194500975
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ja",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3589526769
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"jv",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"jv",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.1994883012
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"jv",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.3942042616
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"jv",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2570478693
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"jv",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5215463463
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"mr",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"mr",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.1828389227
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"mr",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.3179756072
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"mr",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2129586558
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"mr",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4314516197
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pa",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pa",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.252814761
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pa",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4131775231
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pa",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2671823746
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pa",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3966391033
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pt",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pt",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pt",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2348823133
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pt",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4500923911
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pt",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4182790857
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"pt",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.649967582
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ru",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ru",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ru",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.1839314203
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ru",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.4234191674
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ru",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3489148579
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ru",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5567945257
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"sw",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"sw",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"sw",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2041105012
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"sw",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.3657796945
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"sw",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3130978532
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"sw",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5950962977
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"te",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"te",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"te",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.1200373123
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"te",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.3196364935
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"te",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.1814754432
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"te",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3157059838
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ur",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ur",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.181272453
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ur",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.386776605
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ur",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2603415771
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"ur",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4241396601
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"vi",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"vi",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"vi",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.1705385375
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"vi",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.3747437419
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"vi",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.257036702
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"vi",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4598854693
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"zh",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.1
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"zh",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.0
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"zh",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2137844239
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"zh",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5028557922
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"zh",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3129908127
-  },
-  {
-    "model":"deepseek\/deepseek-r1",
-    "bcp_47":"zh",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3710290799
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ar",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ar",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ar",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3412745465
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ar",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5826925715
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ar",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.390058702
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ar",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.597901367
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"bn",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"bn",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"bn",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2988263355
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"bn",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5680625724
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"bn",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3895386095
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"bn",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5693830083
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"de",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"de",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"de",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.4245142301
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"de",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.6390046108
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"de",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4937713215
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"de",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6856946146
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"en",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"en",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"en",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.5700014681
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"en",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.7211557733
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"en",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.6711231356
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"en",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.8279616884
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"es",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"es",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"es",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3344072959
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"es",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5751886204
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"es",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3905772718
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"es",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6166676981
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"fr",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"fr",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"fr",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3411643102
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"fr",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5823324237
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"fr",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.549646027
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"fr",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.7334440232
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"hi",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"hi",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"hi",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.4434404142
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"hi",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.6607832834
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"hi",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4308263411
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"hi",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6353494719
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"id",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"id",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"id",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3626330278
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"id",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5869542517
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"id",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4368014195
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"id",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6964617832
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ja",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ja",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ja",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3331002174
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ja",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.579659503
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ja",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.336969824
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ja",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4913494739
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"jv",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"jv",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3700621486
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"jv",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5837249923
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"jv",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3088767184
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"jv",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5884562937
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"mr",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"mr",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3338185652
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"mr",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5787491818
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"mr",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2806196555
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"mr",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4516145469
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pa",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pa",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.4372599799
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pa",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.658993109
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pa",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4174871385
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pa",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5931595705
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pt",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pt",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pt",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3198419121
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pt",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.578067895
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pt",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.5056809967
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"pt",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6966172136
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ru",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ru",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ru",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.2839962776
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ru",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5461994333
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ru",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.4589390255
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ru",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6469214364
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"sw",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"sw",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"sw",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.4043048116
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"sw",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5944297519
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"sw",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.499303533
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"sw",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.7212747243
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"te",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"te",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.7
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"te",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.4243816819
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"te",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.6336752528
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"te",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3644641049
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"te",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.5789291672
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ur",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":1.0
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ur",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.31244285
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ur",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5630479496
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ur",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.2911146353
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"ur",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.4826415387
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"vi",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"vi",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"vi",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3436662566
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"vi",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.59026429
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"vi",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.447007323
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"vi",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.6624838094
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"zh",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"zh",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.8
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"zh",
-    "task":"translation_from",
-    "metric":"bleu",
-    "score":0.3075560511
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"zh",
-    "task":"translation_from",
-    "metric":"chrf",
-    "score":0.5721434339
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"zh",
-    "task":"translation_to",
-    "metric":"bleu",
-    "score":0.3628421643
-  },
-  {
-    "model":"google\/gemini-2.0-flash-001",
-    "bcp_47":"zh",
-    "task":"translation_to",
-    "metric":"chrf",
-    "score":0.3921364269
-  },
-  {
-    "model":"google\/gemini-2.0-flash-lite-001",
-    "bcp_47":"ar",
-    "task":"classification",
-    "metric":"accuracy",
-    "score":0.9
-  },
-  {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2723646085
+    "score":0.2238237549
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.523730035
+    "score":0.4534366926
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3878873306
+    "score":0.2689055687
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.592059186
+    "score":0.4149416248
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.2
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2672489266
+    "score":0.2443635406
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4952818431
+    "score":0.4676159664
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3639249804
+    "score":0.2223182846
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5538942616
+    "score":0.3685961254
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"de",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.7
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"de",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3382313883
+    "score":0.2633055293
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5705047774
+    "score":0.451378667
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5354923841
+    "score":0.2576565152
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7125417889
+    "score":0.3874723625
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
+    "score":0.1
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4617498931
+    "score":0.3974650186
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6321211549
+    "score":0.5373829936
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5951204739
+    "score":0.5868227988
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7613397345
+    "score":0.7574063883
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3547349164
+    "score":0.2163623393
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5728186386
+    "score":0.4705179867
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3576086442
+    "score":0.2458237388
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6061019948
+    "score":0.426538099
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"fr",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.7
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"fr",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2978444639
+    "score":0.3182982487
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5625976718
+    "score":0.5389072956
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5244850062
+    "score":0.47314841
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7081676298
+    "score":0.6884710951
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.2
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4286739426
+    "score":0.2684001499
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6359453866
+    "score":0.416230929
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4344801023
+    "score":0.2827365983
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6474630132
+    "score":0.5128198247
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.2
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3126497326
+    "score":0.223787985
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5733116043
+    "score":0.485186041
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3882402101
+    "score":0.3571004344
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6624914478
+    "score":0.6398491182
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.2
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ja",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ja",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.289402493
+    "score":0.1711796281
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5452343766
+    "score":0.4108641598
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1879084772
+    "score":0.2194500975
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3787664659
+    "score":0.3589526769
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3561823456
+    "score":0.1994883012
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5619277442
+    "score":0.3942042616
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2609272138
+    "score":0.2570478693
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5344454302
+    "score":0.5215463463
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2906692433
+    "score":0.1828389227
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5058194876
+    "score":0.3179756072
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2450137243
+    "score":0.2129586558
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4736891146
+    "score":0.4314516197
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3792889323
+    "score":0.252814761
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6001893022
+    "score":0.4131775231
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4267930078
+    "score":0.2671823746
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.583867208
+    "score":0.3966391033
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3275326122
+    "score":0.2348823133
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5672682921
+    "score":0.4500923911
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4390630519
+    "score":0.4182790857
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6588223005
+    "score":0.649967582
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.266439807
+    "score":0.1839314203
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5101737799
+    "score":0.4234191674
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4085478092
+    "score":0.3489148579
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5658655698
+    "score":0.5567945257
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"sw",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"sw",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2763965367
+    "score":0.2041105012
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.520411275
+    "score":0.3657796945
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4350912598
+    "score":0.3130978532
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6853284539
+    "score":0.5950962977
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3588325355
+    "score":0.1200373123
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5889091872
+    "score":0.3196364935
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3231474889
+    "score":0.1814754432
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5470305035
+    "score":0.3157059838
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2859928961
+    "score":0.181272453
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5295533045
+    "score":0.386776605
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3010091175
+    "score":0.2603415771
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.493353195
+    "score":0.4241396601
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.0
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3170290484
+    "score":0.1705385375
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5491010642
+    "score":0.3747437419
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4032471641
+    "score":0.257036702
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6394566771
+    "score":0.4598854693
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.1
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"zh",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"zh",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2557280993
+    "score":0.2137844239
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5212143675
+    "score":0.5028557922
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2610248692
+    "score":0.3129908127
   },
   {
-    "model":"google\/gemini-2.0-flash-lite-001",
+    "model":"deepseek\/deepseek-r1",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.314329989
+    "score":0.3710290799
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3462011916
+    "score":0.3412745465
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5813984303
+    "score":0.5826925715
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4319399394
+    "score":0.390058702
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6269389132
+    "score":0.597901367
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3332105938
+    "score":0.2988263355
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5788748413
+    "score":0.5680625724
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3925309816
+    "score":0.3895386095
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5671456285
+    "score":0.5693830083
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"de",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"de",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4080538745
+    "score":0.4245142301
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6278026445
+    "score":0.6390046108
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4940498755
+    "score":0.4937713215
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6915113583
+    "score":0.6856946146
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"en",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"en",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.5780592594
+    "score":0.5700014681
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.727062887
+    "score":0.7211557733
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.725042796
+    "score":0.6711231356
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.868453404
+    "score":0.8279616884
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3228577771
+    "score":0.3344072959
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5691176688
+    "score":0.5751886204
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4211069488
+    "score":0.3905772718
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6329026504
+    "score":0.6166676981
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"fr",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"fr",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3683003934
+    "score":0.3411643102
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5961800274
+    "score":0.5823324237
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5899042948
+    "score":0.549646027
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7506595836
+    "score":0.7334440232
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4661333335
+    "score":0.4434404142
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6599650679
+    "score":0.6607832834
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4319599559
+    "score":0.4308263411
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6493733048
+    "score":0.6353494719
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3507436052
+    "score":0.3626330278
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5856872774
+    "score":0.5869542517
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.387557065
+    "score":0.4368014195
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6730969345
+    "score":0.6964617832
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3590987818
+    "score":0.3331002174
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5958999814
+    "score":0.579659503
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4151803591
+    "score":0.336969824
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5319028037
+    "score":0.4913494739
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3793672083
+    "score":0.3700621486
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6054760312
+    "score":0.5837249923
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3161702144
+    "score":0.3088767184
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5956895972
+    "score":0.5884562937
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2888720948
+    "score":0.3338185652
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5411662197
+    "score":0.5787491818
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2820351359
+    "score":0.2806196555
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5017455952
+    "score":0.4516145469
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4833653569
+    "score":0.4372599799
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6877252787
+    "score":0.658993109
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4959043786
+    "score":0.4174871385
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.631939007
+    "score":0.5931595705
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4185968559
+    "score":0.3198419121
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6194298291
+    "score":0.578067895
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5066913819
+    "score":0.5056809967
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7045563325
+    "score":0.6966172136
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ru",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ru",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2835606416
+    "score":0.2839962776
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5525332546
+    "score":0.5461994333
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5294387768
+    "score":0.4589390255
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6814959444
+    "score":0.6469214364
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4452506057
+    "score":0.4043048116
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6299322033
+    "score":0.5944297519
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4773237879
+    "score":0.499303533
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6983325706
+    "score":0.7212747243
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4838983303
+    "score":0.4243816819
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6878651593
+    "score":0.6336752528
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4225689448
+    "score":0.3644641049
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6171822766
+    "score":0.5789291672
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3194163699
+    "score":0.31244285
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5758075077
+    "score":0.5630479496
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2609063103
+    "score":0.2911146353
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4351160146
+    "score":0.4826415387
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4227968054
+    "score":0.3436662566
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6326522416
+    "score":0.59026429
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4693432911
+    "score":0.447007323
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6659480306
+    "score":0.6624838094
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2632355471
+    "score":0.3075560511
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5527183044
+    "score":0.5721434339
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3128970722
+    "score":0.3628421643
   },
   {
-    "model":"google\/gemini-2.5-flash-preview",
+    "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3699956681
+    "score":0.3921364269
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3509446122
+    "score":0.2723646085
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5872861306
+    "score":0.523730035
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4369226272
+    "score":0.3878873306
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6357392299
+    "score":0.592059186
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2928560114
+    "score":0.2672489266
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5373877759
+    "score":0.4952818431
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.371713483
+    "score":0.3639249804
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5467648432
+    "score":0.5538942616
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"de",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"de",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3827187056
+    "score":0.3382313883
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6092943987
+    "score":0.5705047774
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5181328066
+    "score":0.5354923841
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6981190119
+    "score":0.7125417889
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.5942868447
+    "score":0.4617498931
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.7298379077
+    "score":0.6321211549
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.7309888475
+    "score":0.5951204739
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.8534239972
+    "score":0.7613397345
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"es",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"es",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3497891701
+    "score":0.3547349164
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5889817614
+    "score":0.5728186386
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.395885417
+    "score":0.3576086442
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6273835234
+    "score":0.6061019948
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.7
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3737227872
+    "score":0.2978444639
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6039484183
+    "score":0.5625976718
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5555743798
+    "score":0.5244850062
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7229519037
+    "score":0.7081676298
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4387215588
+    "score":0.4286739426
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6443095169
+    "score":0.6359453866
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4255607348
+    "score":0.4344801023
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6457270128
+    "score":0.6474630132
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3789722313
+    "score":0.3126497326
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5973074889
+    "score":0.5733116043
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.406320567
+    "score":0.3882402101
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6828299565
+    "score":0.6624914478
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3423388267
+    "score":0.289402493
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5884893314
+    "score":0.5452343766
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3740472584
+    "score":0.1879084772
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4832664839
+    "score":0.3787664659
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3540174328
+    "score":0.3561823456
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5513315973
+    "score":0.5619277442
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.329885986
+    "score":0.2609272138
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5805485408
+    "score":0.5344454302
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3055522766
+    "score":0.2906692433
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5435968357
+    "score":0.5058194876
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.257460039
+    "score":0.2450137243
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4845716024
+    "score":0.4736891146
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4182659606
+    "score":0.3792889323
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6696655487
+    "score":0.6001893022
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4665320226
+    "score":0.4267930078
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6280330147
+    "score":0.583867208
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3485890022
+    "score":0.3275326122
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.574915747
+    "score":0.5672682921
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4760104707
+    "score":0.4390630519
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6891420805
+    "score":0.6588223005
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2893557086
+    "score":0.266439807
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5514512546
+    "score":0.5101737799
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4910352633
+    "score":0.4085478092
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6636379715
+    "score":0.5658655698
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.397583037
+    "score":0.2763965367
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6014699575
+    "score":0.520411275
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4540021563
+    "score":0.4350912598
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7002830052
+    "score":0.6853284539
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"te",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"te",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4025505277
+    "score":0.3528600199
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6339036498
+    "score":0.5850279702
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4166086163
+    "score":0.3231474889
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6171953291
+    "score":0.5470305035
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3212777323
+    "score":0.2859928961
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5735084601
+    "score":0.5295533045
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3027958989
+    "score":0.3010091175
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4912965943
+    "score":0.493353195
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
     "score":1.0
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3471784526
+    "score":0.3170290484
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5719753053
+    "score":0.5491010642
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4601231177
+    "score":0.4032471641
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6644848789
+    "score":0.6394566771
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2989639846
+    "score":0.2557280993
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5646599732
+    "score":0.5212143675
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2915310914
+    "score":0.2610248692
   },
   {
-    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3358004159
+    "score":0.314329989
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2747843596
+    "score":0.3462011916
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5519960681
+    "score":0.5813984303
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2863967069
+    "score":0.4319399394
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5318173199
+    "score":0.6269389132
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.271237739
+    "score":0.3332105938
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5173954387
+    "score":0.5788748413
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3067537945
+    "score":0.3925309816
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5194482945
+    "score":0.5671456285
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3189602129
+    "score":0.4080538745
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5548503533
+    "score":0.6278026445
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5229096392
+    "score":0.4940498755
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7023434262
+    "score":0.6915113583
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"en",
-    "task":"mmlu",
+    "task":"mgsm",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"en",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3889146477
+    "score":0.5780592594
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.628092835
+    "score":0.727062887
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4660772497
+    "score":0.725042796
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7280386297
+    "score":0.868453404
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3593767686
+    "score":0.3228577771
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5668073679
+    "score":0.5691176688
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3662275621
+    "score":0.4211069488
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6102640711
+    "score":0.6329026504
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3180384008
+    "score":0.3683003934
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5571267732
+    "score":0.5961800274
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5778354146
+    "score":0.5899042948
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7490356238
+    "score":0.7506595836
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3722751955
+    "score":0.4661333335
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6207213131
+    "score":0.6599650679
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3378499277
+    "score":0.4319599559
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.554090013
+    "score":0.6493733048
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3583744222
+    "score":0.3507436052
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5732194975
+    "score":0.5856872774
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3364664006
+    "score":0.387557065
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6438910651
+    "score":0.6730969345
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2581140706
+    "score":0.3590987818
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5395853617
+    "score":0.5958999814
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3298839393
+    "score":0.4151803591
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4471547552
+    "score":0.5319028037
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3177915441
+    "score":0.3793672083
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5387853038
+    "score":0.6054760312
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2549228547
+    "score":0.3161702144
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5322440265
+    "score":0.5956895972
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2370074805
+    "score":0.2888720948
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4804215458
+    "score":0.5411662197
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2399769139
+    "score":0.2820351359
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4726429935
+    "score":0.5017455952
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4180718844
+    "score":0.4833653569
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6426219278
+    "score":0.6877252787
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3938693136
+    "score":0.4959043786
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5573992167
+    "score":0.631939007
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3537745123
+    "score":0.4185968559
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5904429929
+    "score":0.6194298291
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4588664196
+    "score":0.5066913819
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6844540285
+    "score":0.7045563325
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2449777422
+    "score":0.2835606416
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5268764903
+    "score":0.5525332546
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3789708434
+    "score":0.5294387768
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5790333031
+    "score":0.6814959444
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3005035588
+    "score":0.4452506057
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5210660172
+    "score":0.6299322033
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3401968092
+    "score":0.4773237879
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6217197146
+    "score":0.6983325706
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3330093484
+    "score":0.4838983303
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5694168709
+    "score":0.6878651593
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2667299931
+    "score":0.4225689448
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5208033902
+    "score":0.6171822766
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.322448107
+    "score":0.3194163699
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5887654616
+    "score":0.5758075077
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2148139783
+    "score":0.2609063103
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4226865444
+    "score":0.4351160146
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2745000434
+    "score":0.4227968054
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5206422805
+    "score":0.6326522416
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4261790941
+    "score":0.4693432911
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6358462464
+    "score":0.6659480306
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2201641871
+    "score":0.2632355471
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5051068628
+    "score":0.5527183044
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2508351517
+    "score":0.3128970722
   },
   {
-    "model":"google\/gemini-flash-1.5",
+    "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3110461024
+    "score":0.3699956681
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.260683336
+    "score":0.3509446122
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5383651277
+    "score":0.5872861306
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2962406565
+    "score":0.4369226272
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5190026627
+    "score":0.6357392299
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2183929994
+    "score":0.2928560114
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4877941086
+    "score":0.5373877759
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2700916391
+    "score":0.371713483
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5150577414
+    "score":0.5467648432
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3440655166
+    "score":0.3827187056
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.563902418
+    "score":0.6092943987
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4305522274
+    "score":0.5181328066
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6477508732
+    "score":0.6981190119
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4113463435
+    "score":0.5942868447
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6260248317
+    "score":0.7298379077
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5310035709
+    "score":0.7309888475
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7595845064
+    "score":0.8534239972
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3031284355
+    "score":0.3497891701
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5241309352
+    "score":0.5889817614
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3289699508
+    "score":0.395885417
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5811203167
+    "score":0.6273835234
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3349110908
+    "score":0.3737227872
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.564806297
+    "score":0.6039484183
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4574014191
+    "score":0.5555743798
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6626552528
+    "score":0.7229519037
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3184042229
+    "score":0.4387215588
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5712698408
+    "score":0.6443095169
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3659029431
+    "score":0.4255607348
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5796195236
+    "score":0.6457270128
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2301748885
+    "score":0.3789722313
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4682741896
+    "score":0.5973074889
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3346592082
+    "score":0.406320567
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6514874668
+    "score":0.6828299565
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2664966821
+    "score":0.3423388267
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5386982677
+    "score":0.5884893314
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2651365589
+    "score":0.3740472584
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.409095006
+    "score":0.4832664839
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2326358655
+    "score":0.3540174328
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4815897231
+    "score":0.5513315973
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2187928356
+    "score":0.329885986
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4896578943
+    "score":0.5805485408
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2256246926
+    "score":0.3055522766
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4703189943
+    "score":0.5435968357
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2254492518
+    "score":0.257460039
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.447826525
+    "score":0.4845716024
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3706063992
+    "score":0.4182659606
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6167676482
+    "score":0.6696655487
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4234596823
+    "score":0.4665320226
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5629443923
+    "score":0.6280330147
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2971403532
+    "score":0.3485890022
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5321068893
+    "score":0.574915747
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4364286549
+    "score":0.4760104707
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6509885745
+    "score":0.6891420805
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2300270544
+    "score":0.2893557086
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4839384065
+    "score":0.5514512546
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4236492288
+    "score":0.4910352633
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6116207052
+    "score":0.6636379715
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2501435914
+    "score":0.397583037
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5088299265
+    "score":0.6014699575
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.262372343
+    "score":0.4540021563
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5806899403
+    "score":0.7002830052
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3568851036
+    "score":0.4025505277
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5825326367
+    "score":0.6339036498
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3112091725
+    "score":0.4166086163
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5431414206
+    "score":0.6171953291
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2402951661
+    "score":0.3212777323
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5033005385
+    "score":0.5735084601
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2302239803
+    "score":0.3027958989
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4066956434
+    "score":0.4912965943
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2537752957
+    "score":0.3471784526
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5073147534
+    "score":0.5719753053
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3583753747
+    "score":0.4601231177
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6253917282
+    "score":0.6644848789
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.249810194
+    "score":0.2989639846
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5151255506
+    "score":0.5646599732
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1854861198
+    "score":0.2915310914
   },
   {
-    "model":"google\/gemini-flash-1.5-8b",
+    "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2463326959
+    "score":0.3358004159
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":1.0
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2170512634
+    "score":0.2747843596
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4780493622
+    "score":0.5519960681
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2844387275
+    "score":0.2863967069
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4979841291
+    "score":0.5318173199
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.6
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1842659234
+    "score":0.271237739
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4331372452
+    "score":0.5173954387
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2644024965
+    "score":0.3067537945
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.462996411
+    "score":0.5194482945
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.6
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2470808909
+    "score":0.3189602129
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4959472303
+    "score":0.5548503533
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3105509871
+    "score":0.5229096392
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5118330277
+    "score":0.7023434262
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.8
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4810009804
+    "score":0.3889146477
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6244199222
+    "score":0.628092835
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.6112171119
+    "score":0.4660772497
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.8046069838
+    "score":0.7280386297
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.1
+    "score":0.8
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1990314823
+    "score":0.3593767686
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4557184175
+    "score":0.5668073679
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3019092795
+    "score":0.3662275621
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5373780713
+    "score":0.6102640711
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.8
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1741359314
+    "score":0.3180384008
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4175910256
+    "score":0.5571267732
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4029915732
+    "score":0.5778354146
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.618701934
+    "score":0.7490356238
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.1
+    "score":1.0
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3016091966
+    "score":0.3722751955
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5281403748
+    "score":0.6207213131
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2736270122
+    "score":0.3378499277
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5159089396
+    "score":0.554090013
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1930481943
+    "score":0.3583744222
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.420733839
+    "score":0.5732194975
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.344885086
+    "score":0.3364664006
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6497564532
+    "score":0.6438910651
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.8
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2254647544
+    "score":0.2581140706
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4810564434
+    "score":0.5395853617
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ja",
+    "task":"translation_to",
+    "metric":"bleu",
+    "score":0.3298839393
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ja",
     "task":"translation_to",
+    "metric":"chrf",
+    "score":0.4471547552
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"jv",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"jv",
+    "task":"translation_from",
+    "metric":"bleu",
+    "score":0.3177915441
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"jv",
+    "task":"translation_from",
+    "metric":"chrf",
+    "score":0.5387853038
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"jv",
+    "task":"translation_to",
     "metric":"bleu",
-    "score":0.1874016498
+    "score":0.2549228547
   },
   {
-    "model":"google\/gemma-2-9b-it",
-    "bcp_47":"ja",
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2916976073
+    "score":0.5322440265
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1616328428
+    "score":0.2370074805
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.414450627
+    "score":0.4804215458
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1123956753
+    "score":0.2399769139
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3542849129
+    "score":0.4726429935
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.34113337
+    "score":0.4180718844
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5187033084
+    "score":0.6426219278
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3611338165
+    "score":0.3938693136
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5059192583
+    "score":0.5573992167
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.1
+    "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2429635608
+    "score":0.3537745123
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.473106986
+    "score":0.5904429929
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2843181347
+    "score":0.4588664196
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5735245279
+    "score":0.6844540285
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":1.0
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1697780011
+    "score":0.2449777422
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4361163813
+    "score":0.5268764903
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3672218817
+    "score":0.3789708434
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5564712622
+    "score":0.5790333031
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.8
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.188207488
+    "score":0.3005035588
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4728711823
+    "score":0.5210660172
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.27482714
+    "score":0.3401968092
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5603409861
+    "score":0.6217197146
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.6
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3095013641
+    "score":0.3330093484
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5252619118
+    "score":0.5694168709
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1798656271
+    "score":0.257812168
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4011450412
+    "score":0.5167002436
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.137296648
+    "score":0.322448107
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.341408254
+    "score":0.5887654616
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0595669677
+    "score":0.2148139783
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.201487497
+    "score":0.4226865444
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"vi",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"vi",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"vi",
+    "task":"translation_from",
+    "metric":"bleu",
+    "score":0.2745000434
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"vi",
+    "task":"translation_from",
+    "metric":"chrf",
+    "score":0.5206422805
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"vi",
+    "task":"translation_to",
+    "metric":"bleu",
+    "score":0.4261790941
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"vi",
+    "task":"translation_to",
+    "metric":"chrf",
+    "score":0.6358462464
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.1
+    "score":0.8
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1922618764
+    "score":0.2201641871
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4517742384
+    "score":0.5051068628
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2125812397
+    "score":0.2508351517
   },
   {
-    "model":"google\/gemma-2-9b-it",
+    "model":"google\/gemini-flash-1.5",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2689557345
+    "score":0.3110461024
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.4
+    "score":0.7
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2891206499
+    "score":0.260683336
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5438550217
+    "score":0.5383651277
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3184721364
+    "score":0.2962406565
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5483731849
+    "score":0.5190026627
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.6
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2907230812
+    "score":0.2183929994
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5148223626
+    "score":0.4877941086
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3473636391
+    "score":0.2700916391
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5442574441
+    "score":0.5150577414
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.8
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3577876868
+    "score":0.3440655166
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5943423055
+    "score":0.563902418
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5344280565
+    "score":0.4305522274
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7084649844
+    "score":0.6477508732
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.452427177
+    "score":0.4113463435
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6525566656
+    "score":0.6260248317
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5467976399
+    "score":0.5310035709
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7780833183
+    "score":0.7595845064
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.6
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3554331718
+    "score":0.3031284355
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5997743406
+    "score":0.5241309352
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3847830842
+    "score":0.3289699508
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6191109047
+    "score":0.5811203167
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.8
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2846092378
+    "score":0.3349110908
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5655970541
+    "score":0.564806297
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5192984544
+    "score":0.4574014191
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7020040834
+    "score":0.6626552528
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.6
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3520691191
+    "score":0.3184042229
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6035990708
+    "score":0.5712698408
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3987037224
+    "score":0.3659029431
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6195037668
+    "score":0.5796195236
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.5
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3377417704
+    "score":0.2301748885
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5674360496
+    "score":0.4682741896
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3534620252
+    "score":0.3346592082
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6680177029
+    "score":0.6514874668
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.4
+    "score":0.7
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2481856237
+    "score":0.2664966821
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5180749152
+    "score":0.5386982677
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.242508046
+    "score":0.2651365589
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4046420215
+    "score":0.409095006
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3019627022
+    "score":0.2326358655
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5133980923
+    "score":0.4815897231
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2316517545
+    "score":0.2187928356
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5189963647
+    "score":0.4896578943
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3395095603
+    "score":0.2256246926
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5877742809
+    "score":0.4703189943
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2222923122
+    "score":0.2254492518
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4572688692
+    "score":0.447826525
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4164890636
+    "score":0.3706063992
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6375470445
+    "score":0.6167676482
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4166823661
+    "score":0.4234596823
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.575314128
+    "score":0.5629443923
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.340245547
+    "score":0.2971403532
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5681284927
+    "score":0.5321068893
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4702737577
+    "score":0.4364286549
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6903236014
+    "score":0.6509885745
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.8
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2862936285
+    "score":0.2300270544
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5264436928
+    "score":0.4839384065
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4800957551
+    "score":0.4236492288
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6618495803
+    "score":0.6116207052
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.268709657
+    "score":0.2501435914
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5306834056
+    "score":0.5088299265
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3395981599
+    "score":0.262372343
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6130756934
+    "score":0.5806899403
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.2
+    "score":0.6
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.279412176
+    "score":0.3568851036
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5419647491
+    "score":0.5825326367
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.259866454
+    "score":0.3112091725
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.48543634
+    "score":0.5431414206
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2595582459
+    "score":0.2402951661
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5043992681
+    "score":0.5033005385
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2142625601
+    "score":0.2302239803
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.429749938
+    "score":0.4066956434
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.4
+    "score":0.6
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2279880384
+    "score":0.2537752957
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4835933272
+    "score":0.5073147534
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3478085621
+    "score":0.3583753747
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5968604742
+    "score":0.6253917282
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.9
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1900086584
+    "score":0.249810194
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4895930442
+    "score":0.5151255506
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2395565562
+    "score":0.1854861198
   },
   {
-    "model":"google\/gemma-3-27b-it",
+    "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3237759485
+    "score":0.2463326959
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.5
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.4
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0837265107
+    "score":0.2891206499
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3012065838
+    "score":0.5438550217
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1120670716
+    "score":0.3184721364
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2771296913
+    "score":0.5483731849
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.4
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.3
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.021858254
+    "score":0.2907230812
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2327499821
+    "score":0.5148223626
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0253088472
+    "score":0.3473636391
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.1918662187
+    "score":0.5442574441
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.4
+    "score":0.7
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1784955678
+    "score":0.3577876868
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4008799371
+    "score":0.5943423055
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1782663616
+    "score":0.5344280565
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4070825897
+    "score":0.7084649844
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.6
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2797104835
+    "score":0.452427177
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4601471921
+    "score":0.6525566656
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3392404298
+    "score":0.5467976399
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5518674496
+    "score":0.7780833183
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.4
+    "score":0.6
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1682956348
+    "score":0.3554331718
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3705606944
+    "score":0.5997743406
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1595479626
+    "score":0.3847830842
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4040956812
+    "score":0.6191109047
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.3
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1362099506
+    "score":0.2846092378
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4039646029
+    "score":0.5655970541
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2151785904
+    "score":0.5192984544
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4309035319
+    "score":0.7020040834
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.6
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.2
+    "score":0.5
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1658307051
+    "score":0.3520691191
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3676480008
+    "score":0.6035990708
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1711676323
+    "score":0.3987037224
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.360371738
+    "score":0.6195037668
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.9
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.3
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1399603895
+    "score":0.3377417704
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3604113675
+    "score":0.5674360496
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1312450274
+    "score":0.3534620252
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3988827234
+    "score":0.6680177029
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.7
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.1
+    "score":0.4
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1222242366
+    "score":0.2481856237
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3415186168
+    "score":0.5180749152
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1607198845
+    "score":0.242508046
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2706371796
+    "score":0.4046420215
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.9
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0550016522
+    "score":0.3019627022
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2468917982
+    "score":0.5133980923
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.017690474
+    "score":0.2316517545
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2536001746
+    "score":0.5189963647
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.4
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0411296617
+    "score":0.3395095603
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2442292695
+    "score":0.5877742809
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0482775033
+    "score":0.2222923122
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2211003771
+    "score":0.4572688692
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.0
+    "score":0.9
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2292007848
+    "score":0.4164890636
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3716655897
+    "score":0.6375470445
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2126702079
+    "score":0.4166823661
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3248815955
+    "score":0.575314128
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.7
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.4
+    "score":0.6
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1013445398
+    "score":0.340245547
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3063125264
+    "score":0.5681284927
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2231857524
+    "score":0.4702737577
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4401829864
+    "score":0.6903236014
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
     "score":0.8
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.3
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1005466956
+    "score":0.2862936285
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3356518748
+    "score":0.5264436928
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1363185356
+    "score":0.4800957551
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3769111636
+    "score":0.6618495803
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.5
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.6
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0510060878
+    "score":0.268709657
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2132741272
+    "score":0.5306834056
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0089146903
+    "score":0.3395981599
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2314109768
+    "score":0.6130756934
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.2
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.1
+    "score":0.2
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1096331511
+    "score":0.2787922254
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2888090685
+    "score":0.5420797212
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1206114883
+    "score":0.259866454
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2597106436
+    "score":0.48543634
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.9
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.0891185343
+    "score":0.2595582459
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.281793335
+    "score":0.5043992681
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.0799576366
+    "score":0.2142625601
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2661903898
+    "score":0.429749938
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.2
+    "score":0.4
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1389065496
+    "score":0.2279880384
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3708687542
+    "score":0.4835933272
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1661316612
+    "score":0.3478085621
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.354399593
+    "score":0.5968604742
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.7
+    "score":1.0
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.5
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1133138775
+    "score":0.1900086584
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3754190494
+    "score":0.4895930442
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1672636279
+    "score":0.2395565562
   },
   {
-    "model":"gryphe\/mythomax-l2-13b",
+    "model":"google\/gemma-3-27b-it",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.232325354
+    "score":0.3237759485
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.5
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.4
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2609114367
+    "score":0.0837265107
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5167379854
+    "score":0.3012065838
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1932802581
+    "score":0.1120670716
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4648835751
+    "score":0.2771296913
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.4
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.3
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1938367121
+    "score":0.021858254
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4674774016
+    "score":0.2327499821
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2073541352
+    "score":0.0253088472
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4438396219
+    "score":0.1918662187
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.4
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2685056004
+    "score":0.1784955678
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.506818165
+    "score":0.4008799371
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3801657831
+    "score":0.1782663616
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6056477234
+    "score":0.4070825897
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2216031518
+    "score":0.2797104835
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4453474399
+    "score":0.4601471921
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3322887566
+    "score":0.3392404298
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5824514758
+    "score":0.5518674496
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.4
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1640783778
+    "score":0.1682956348
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.434749516
+    "score":0.3705606944
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3126220052
+    "score":0.1595479626
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6038883227
+    "score":0.4040956812
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"fr",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"fr",
-    "task":"classification",
+    "task":"mgsm",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.0
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.5
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2261588318
+    "score":0.1362099506
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4944996319
+    "score":0.4039646029
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4626001556
+    "score":0.2151785904
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6614963779
+    "score":0.4309035319
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.6
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.2
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3316295853
+    "score":0.1658307051
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5594083443
+    "score":0.3676480008
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.365274938
+    "score":0.1711676323
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.588222932
+    "score":0.360371738
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.3
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2236143729
+    "score":0.1399603895
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4912948296
+    "score":0.3604113675
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2910526755
+    "score":0.1312450274
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6005590773
+    "score":0.3988827234
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ja",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.1
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2131020144
+    "score":0.1222242366
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4845704057
+    "score":0.3415186168
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1446650781
+    "score":0.1607198845
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2292145443
+    "score":0.2706371796
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.6
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1840709267
+    "score":0.0550016522
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4042090141
+    "score":0.2468917982
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1902389614
+    "score":0.017690474
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4796942089
+    "score":0.2536001746
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.4
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2152417217
+    "score":0.0411296617
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4578207034
+    "score":0.2442292695
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1654073391
+    "score":0.0482775033
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3941079443
+    "score":0.2211003771
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.0
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1891048622
+    "score":0.2292007848
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4233257981
+    "score":0.3716655897
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2086330089
+    "score":0.2126702079
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.364326938
+    "score":0.3248815955
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.4
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2256776552
+    "score":0.1013445398
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4793769886
+    "score":0.3063125264
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3320822339
+    "score":0.2231857524
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6132478102
+    "score":0.4401829864
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.3
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1730354472
+    "score":0.1005466956
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.454951133
+    "score":0.3356518748
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3073459183
+    "score":0.1363185356
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.554890569
+    "score":0.3769111636
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.5
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.3
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1798653655
+    "score":0.0510060878
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4290347271
+    "score":0.2132741272
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1709237193
+    "score":0.0089146903
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.522080463
+    "score":0.2314109768
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.2
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.1
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2507091988
+    "score":0.1096331511
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5067409185
+    "score":0.2888090685
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2998213366
+    "score":0.1206114883
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5112516267
+    "score":0.2597106436
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.5
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.194871957
+    "score":0.0891185343
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4838690709
+    "score":0.281793335
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1908512613
+    "score":0.0799576366
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3855814375
+    "score":0.2661903898
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"vi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.2
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2003733128
+    "score":0.1389065496
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"vi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4613442635
+    "score":0.3708687542
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3326058501
+    "score":0.1661316612
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"vi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5496097026
+    "score":0.354399593
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"zh",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.3
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1869648376
+    "score":0.1133138775
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4664675781
+    "score":0.3754190494
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1696786467
+    "score":0.1672636279
   },
   {
-    "model":"meta-llama\/llama-3-70b-instruct",
+    "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2325846099
+    "score":0.232325354
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ar",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":1.0
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1934231192
+    "score":0.2609114367
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4070435279
+    "score":0.5167379854
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1705154843
+    "score":0.1932802581
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4152198662
+    "score":0.4648835751
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1488002659
+    "score":0.1938367121
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3745722385
+    "score":0.4674774016
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1902972944
+    "score":0.2073541352
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3816229412
+    "score":0.4438396219
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"de",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"de",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1785474661
+    "score":0.2685056004
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4370657832
+    "score":0.506818165
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2574134539
+    "score":0.3801657831
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5270318561
+    "score":0.6056477234
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"en",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"en",
     "task":"mmlu",
     "metric":"accuracy",
     "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3528640776
+    "score":0.2216031518
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5518750903
+    "score":0.4429472312
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4474952526
+    "score":0.3322887566
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6711323725
+    "score":0.5824514758
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"es",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1873108686
+    "score":0.1640783778
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4130794041
+    "score":0.434749516
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.233152279
+    "score":0.3126220052
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5457724764
+    "score":0.6038883227
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"fr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1823885989
+    "score":0.2261588318
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4281583879
+    "score":0.4944996319
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3356151999
+    "score":0.4626001556
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5339127213
+    "score":0.6614963779
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"hi",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2985177739
+    "score":0.3316295853
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5122033781
+    "score":0.5594083443
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2288990675
+    "score":0.3598049012
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4641914485
+    "score":0.5828568956
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"id",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"id",
+    "task":"translation_from",
+    "metric":"bleu",
+    "score":0.2236143729
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"id",
+    "task":"translation_from",
+    "metric":"chrf",
+    "score":0.4912948296
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"id",
+    "task":"translation_to",
+    "metric":"bleu",
+    "score":0.2910526755
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"id",
+    "task":"translation_to",
+    "metric":"chrf",
+    "score":0.6005590773
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ja",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.5
+    "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"id",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1513573791
+    "score":0.2131020144
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"id",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4126710531
+    "score":0.4845704057
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"id",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2153271782
+    "score":0.1446650781
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"id",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5281145734
+    "score":0.2292145443
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"ja",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"jv",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"ja",
-    "task":"mmlu",
-    "metric":"accuracy",
-    "score":0.6
-  },
-  {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"ja",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"jv",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1685400892
+    "score":0.1840709267
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"ja",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"jv",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4152307643
+    "score":0.4042090141
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"ja",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"jv",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.133960866
+    "score":0.1902389614
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"ja",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"jv",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2277435316
+    "score":0.4796942089
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"mr",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1507564245
+    "score":0.2152417217
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3990456964
+    "score":0.4578207034
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1244036875
+    "score":0.1654073391
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3211879294
+    "score":0.3941079443
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2304909323
+    "score":0.1891048622
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4895854606
+    "score":0.4238091524
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2411212371
+    "score":0.2086330089
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4357494606
+    "score":0.364326938
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pt",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.4
+    "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1707431977
+    "score":0.2256776552
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.41325613
+    "score":0.4793769886
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3010483452
+    "score":0.3320822339
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5494427262
+    "score":0.6132478102
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ru",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.6
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1269147932
+    "score":0.1730354472
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4169665509
+    "score":0.454951133
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2000369037
+    "score":0.3073459183
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4178019923
+    "score":0.554890569
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"sw",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.2
+    "score":0.6
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1181160377
+    "score":0.1798653655
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3506298599
+    "score":0.4290347271
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.080193685
+    "score":0.1709237193
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.431036918
+    "score":0.522080463
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"te",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.2
+    "score":0.5
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1477600607
+    "score":0.2507091988
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4004241369
+    "score":0.5067409185
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2008915859
+    "score":0.2998213366
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3876603494
+    "score":0.5112516267
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ur",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1383930526
+    "score":0.194871957
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.3847225326
+    "score":0.4838690709
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1092836008
+    "score":0.1908512613
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2981134702
+    "score":0.3855814375
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
-    "bcp_47":"zh",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"vi",
     "task":"classification",
     "metric":"accuracy",
     "score":0.9
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"vi",
+    "task":"mmlu",
+    "metric":"accuracy",
+    "score":0.8
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"vi",
+    "task":"translation_from",
+    "metric":"bleu",
+    "score":0.2003733128
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"vi",
+    "task":"translation_from",
+    "metric":"chrf",
+    "score":0.4613442635
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"vi",
+    "task":"translation_to",
+    "metric":"bleu",
+    "score":0.3326058501
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"vi",
+    "task":"translation_to",
+    "metric":"chrf",
+    "score":0.5496097026
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"zh",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":1.0
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.3
+    "score":0.8
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1188554471
+    "score":0.1869648376
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.393157353
+    "score":0.4664675781
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1823790731
+    "score":0.1696786467
   },
   {
-    "model":"meta-llama\/llama-3-8b-instruct",
+    "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.24906438
+    "score":0.2325846099
   },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
@@ -12830,6 +10968,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"bn",
@@ -12872,6 +11017,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"de",
@@ -12914,6 +11066,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"en",
@@ -12956,6 +11115,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"es",
@@ -12998,6 +11164,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"fr",
@@ -13124,6 +11297,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ja",
@@ -13313,6 +11493,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ru",
@@ -13355,6 +11542,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"sw",
@@ -13397,6 +11591,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"te",
@@ -13516,6 +11717,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"zh",
@@ -13600,6 +11808,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"bn",
@@ -13642,6 +11857,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"de",
@@ -13684,6 +11906,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"en",
@@ -13696,14 +11925,14 @@
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4135052232
+    "score":0.4124666467
   },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5895253676
+    "score":0.582315662
   },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
@@ -13726,6 +11955,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"es",
@@ -13768,6 +12004,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"fr",
@@ -13836,14 +12079,14 @@
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3052977087
+    "score":0.332068694
   },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5029439287
+    "score":0.5232846835
   },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
@@ -13894,6 +12137,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"ja",
@@ -14083,6 +12333,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"ru",
@@ -14125,6 +12382,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"sw",
@@ -14167,6 +12431,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"te",
@@ -14179,14 +12450,14 @@
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2231414616
+    "score":0.2177874056
   },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4272936469
+    "score":0.4228885108
   },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
@@ -14286,6 +12557,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-8b-instruct",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.1-8b-instruct",
     "bcp_47":"zh",
@@ -14370,6 +12648,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"bn",
@@ -14412,6 +12697,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"de",
@@ -14454,6 +12746,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"en",
@@ -14473,7 +12772,7 @@
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2988653266
+    "score":0.2861009841
   },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
@@ -14496,6 +12795,13 @@
     "metric":"accuracy",
     "score":0.3
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"es",
@@ -14538,6 +12844,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"fr",
@@ -14664,6 +12977,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"ja",
@@ -14781,14 +13101,14 @@
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1995473768
+    "score":0.2072883352
   },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.2187948253
+    "score":0.2238489904
   },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
@@ -14853,6 +13173,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"ru",
@@ -14895,6 +13222,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"sw",
@@ -14937,6 +13271,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"te",
@@ -15056,6 +13397,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"meta-llama\/llama-3.2-1b-instruct",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.2-1b-instruct",
     "bcp_47":"zh",
@@ -15110,28 +13458,28 @@
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2952163543
+    "score":0.2997360932
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5478879017
+    "score":0.552448295
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.285185681
+    "score":0.2837522278
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5169899121
+    "score":0.5150814494
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15140,40 +13488,47 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"bn",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.8
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2565907016
+    "score":0.2646772038
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5135642531
+    "score":0.5157826791
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2990673682
+    "score":0.331255344
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5019041301
+    "score":0.5093924414
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15182,6 +13537,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"de",
@@ -15194,28 +13556,28 @@
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3039463145
+    "score":0.3129672706
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5396134408
+    "score":0.5456543979
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4769793495
+    "score":0.4852409005
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6831000305
+    "score":0.6908376394
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15224,6 +13586,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"en",
@@ -15236,28 +13605,28 @@
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.5363634713
+    "score":0.5510215557
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6717522975
+    "score":0.6909834226
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.6861915605
+    "score":0.6679215449
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.8341405041
+    "score":0.823262947
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15266,40 +13635,47 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"es",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3000860461
+    "score":0.3139331841
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5397261344
+    "score":0.5454623234
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3791924256
+    "score":0.3561290923
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6159217656
+    "score":0.599796306
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15308,47 +13684,54 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"fr",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.6
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2910966189
+    "score":0.2963449909
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5655611568
+    "score":0.5544997379
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5612286605
+    "score":0.5091700689
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7335958119
+    "score":0.690320784
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"hi",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.9
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15362,35 +13745,35 @@
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.37963036
+    "score":0.385731086
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6119011461
+    "score":0.6034244629
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.394637981
+    "score":0.357812342
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6170944347
+    "score":0.5894721809
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"id",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15404,28 +13787,28 @@
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3245838511
+    "score":0.3143630083
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5745245116
+    "score":0.5621460006
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3635605483
+    "score":0.3372718385
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6486137543
+    "score":0.6543894215
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15434,40 +13817,47 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ja",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.6
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2805654401
+    "score":0.284322765
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5274978991
+    "score":0.5304019177
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2667114019
+    "score":0.2212745751
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.378544243
+    "score":0.355950114
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15516,63 +13906,63 @@
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2718900108
+    "score":0.3224419544
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5102541071
+    "score":0.5606155155
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1858815817
+    "score":0.1663847917
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4329686416
+    "score":0.4383662593
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pa",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4659929088
+    "score":0.4519562833
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6558133861
+    "score":0.6328740374
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4108150693
+    "score":0.3918546765
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5726366793
+    "score":0.551810658
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15586,35 +13976,35 @@
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2792136387
+    "score":0.3065957195
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5337079901
+    "score":0.5567333989
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4673129305
+    "score":0.4450524918
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.666759063
+    "score":0.6712742861
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15623,47 +14013,61 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ru",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.8
+    "score":0.7
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2072896517
+    "score":0.2184907643
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4846106947
+    "score":0.4824691404
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3413312048
+    "score":0.3406091079
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5652152557
+    "score":0.5689518318
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"sw",
     "task":"classification",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
+  },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15677,28 +14081,28 @@
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2805870918
+    "score":0.2373393477
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5415843032
+    "score":0.5017230165
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2475190192
+    "score":0.2334583695
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5651818159
+    "score":0.5603415221
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15707,40 +14111,47 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"te",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.7
+    "score":0.8
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3394240055
+    "score":0.3767393472
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5797713804
+    "score":0.5870286691
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3029126381
+    "score":0.3415510311
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5385965131
+    "score":0.5592933672
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15754,28 +14165,28 @@
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2690513839
+    "score":0.252958993
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5193011981
+    "score":0.5053979802
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2616221013
+    "score":0.2339821201
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4396251456
+    "score":0.425306622
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
@@ -15826,6 +14237,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"zh",
@@ -15838,28 +14256,28 @@
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2371072964
+    "score":0.2259180607
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5127276245
+    "score":0.5258178103
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.262487605
+    "score":0.2474954475
   },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3198189866
+    "score":0.2841722148
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -15873,42 +14291,49 @@
     "bcp_47":"ar",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3127757991
+    "score":0.3288891576
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ar",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5649158188
+    "score":0.560493221
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.336779181
+    "score":0.3370289388
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ar",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5507530519
+    "score":0.5616376735
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"bn",
     "task":"classification",
     "metric":"accuracy",
-    "score":1.0
+    "score":0.9
+  },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -15922,28 +14347,28 @@
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3069295097
+    "score":0.3130233588
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"bn",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5513389823
+    "score":0.5569163893
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3483092706
+    "score":0.3322667951
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"bn",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5205648325
+    "score":0.5140240989
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -15952,6 +14377,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"de",
@@ -15964,28 +14396,28 @@
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4025145053
+    "score":0.386239845
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"de",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6178207148
+    "score":0.6110300223
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4799077679
+    "score":0.4508475568
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"de",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6743939478
+    "score":0.6653203029
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -15994,6 +14426,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"en",
@@ -16006,28 +14445,28 @@
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.515872136
+    "score":0.5649266234
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6957157301
+    "score":0.7240609445
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.6059225158
+    "score":0.6160790992
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"en",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.7832269342
+    "score":0.8133281991
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16036,6 +14475,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"es",
@@ -16048,28 +14494,28 @@
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3316519325
+    "score":0.3152773331
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"es",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5678715692
+    "score":0.5548531112
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3488467992
+    "score":0.3335962816
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"es",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.590893987
+    "score":0.5859881472
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16078,6 +14524,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"fr",
@@ -16090,28 +14543,28 @@
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3433334073
+    "score":0.3174175523
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"fr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5948101631
+    "score":0.5682518332
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.5065668297
+    "score":0.5173973527
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"fr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6980371576
+    "score":0.7056428374
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16132,28 +14585,28 @@
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3914498782
+    "score":0.3796955055
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6153407735
+    "score":0.6056742688
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.355440696
+    "score":0.382370623
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5800922474
+    "score":0.5922578575
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16167,35 +14620,35 @@
     "bcp_47":"id",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2363527189
+    "score":0.2523550022
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"id",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5367331703
+    "score":0.5458277736
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3735236798
+    "score":0.3774331947
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"id",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6426286119
+    "score":0.6486159416
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16204,6 +14657,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ja",
@@ -16216,28 +14676,28 @@
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2966381777
+    "score":0.3130983776
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ja",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5565004336
+    "score":0.5536124921
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2985554699
+    "score":0.3266051607
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ja",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.438719005
+    "score":0.4444671407
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16286,28 +14746,28 @@
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3087670334
+    "score":0.3125095049
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"mr",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5771712483
+    "score":0.5535296132
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2457723977
+    "score":0.2275791183
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"mr",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4600595438
+    "score":0.4569915545
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16321,28 +14781,28 @@
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.412130771
+    "score":0.4246522462
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6090479579
+    "score":0.6364852561
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4664487469
+    "score":0.439726575
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pa",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.609059672
+    "score":0.5806732389
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16356,35 +14816,35 @@
     "bcp_47":"pt",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":1.0
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3566326015
+    "score":0.3451568022
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pt",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5934355692
+    "score":0.5845365036
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4571214583
+    "score":0.4582156105
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pt",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6537873967
+    "score":0.6491023878
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16393,6 +14853,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ru",
@@ -16405,28 +14872,28 @@
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2654289532
+    "score":0.2567420946
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ru",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5302105978
+    "score":0.5061211552
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.4272058624
+    "score":0.4025875747
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ru",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6370792063
+    "score":0.6083538055
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16435,6 +14902,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"sw",
@@ -16447,28 +14921,28 @@
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3666782144
+    "score":0.3425500041
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"sw",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5900845247
+    "score":0.5680420054
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2983882308
+    "score":0.2980763573
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"sw",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.6072147424
+    "score":0.5924685945
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16477,6 +14951,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"te",
@@ -16489,28 +14970,28 @@
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3680021687
+    "score":0.3462186566
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6106474057
+    "score":0.6055817314
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3502048845
+    "score":0.3822149946
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5783869442
+    "score":0.5783618359
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16524,28 +15005,28 @@
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2305225364
+    "score":0.253549974
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ur",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4984986116
+    "score":0.5235064606
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2361005323
+    "score":0.2484556869
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ur",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.4656123232
+    "score":0.429442787
   },
   {
     "model":"meta-llama\/llama-4-maverick",
@@ -16596,40 +15077,47 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"zh",
     "task":"mmlu",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.8
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2782569998
+    "score":0.2703289724
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"zh",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5594321689
+    "score":0.5473102513
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2939092084
+    "score":0.317769874
   },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"zh",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3280788158
+    "score":0.3676906362
   },
   {
     "model":"microsoft\/phi-4",
@@ -16680,6 +15168,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"bn",
@@ -16722,6 +15217,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"de",
@@ -16764,6 +15266,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"en",
@@ -16806,6 +15315,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"es",
@@ -16848,6 +15364,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"fr",
@@ -16916,14 +15439,14 @@
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.2425073841
+    "score":0.2369610218
   },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.3817959275
+    "score":0.3765795877
   },
   {
     "model":"microsoft\/phi-4",
@@ -16974,6 +15497,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ja",
@@ -17163,6 +15693,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ru",
@@ -17205,6 +15742,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"sw",
@@ -17247,6 +15791,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"te",
@@ -17259,14 +15810,14 @@
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2322864827
+    "score":0.2374723306
   },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"te",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4651967541
+    "score":0.4675485501
   },
   {
     "model":"microsoft\/phi-4",
@@ -17366,6 +15917,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"zh",
@@ -17450,6 +16008,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"bn",
@@ -17492,6 +16057,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"de",
@@ -17534,6 +16106,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"en",
@@ -17576,6 +16155,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"es",
@@ -17618,6 +16204,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"fr",
@@ -17744,6 +16337,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ja",
@@ -17933,6 +16533,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ru",
@@ -17975,6 +16582,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"sw",
@@ -18017,6 +16631,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"te",
@@ -18136,6 +16757,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"zh",
@@ -18220,6 +16848,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"bn",
@@ -18262,6 +16897,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"de",
@@ -18304,6 +16946,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"en",
@@ -18316,14 +16965,14 @@
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.2963681355
+    "score":0.3166864072
   },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5277177226
+    "score":0.5483508218
   },
   {
     "model":"mistralai\/mistral-nemo",
@@ -18346,6 +16995,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"es",
@@ -18388,6 +17044,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"fr",
@@ -18442,14 +17105,14 @@
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.1878052787
+    "score":0.1271878224
   },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"hi",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.4458139958
+    "score":0.3698932868
   },
   {
     "model":"mistralai\/mistral-nemo",
@@ -18514,6 +17177,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ja",
@@ -18631,14 +17301,14 @@
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3360340722
+    "score":0.3158857772
   },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5243077445
+    "score":0.5109022919
   },
   {
     "model":"mistralai\/mistral-nemo",
@@ -18703,6 +17373,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ru",
@@ -18745,6 +17422,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"sw",
@@ -18787,6 +17471,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"te",
@@ -18813,14 +17504,14 @@
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.1060450795
+    "score":0.2060450795
   },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"te",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.2738653779
+    "score":0.3717773766
   },
   {
     "model":"mistralai\/mistral-nemo",
@@ -18906,6 +17597,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"zh",
@@ -18990,6 +17688,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"bn",
@@ -19032,6 +17737,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"de",
@@ -19074,6 +17786,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"en",
@@ -19086,14 +17805,14 @@
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.4552458759
+    "score":0.4669071745
   },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"en",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.6289172201
+    "score":0.6351301458
   },
   {
     "model":"mistralai\/mistral-saba",
@@ -19116,6 +17835,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"es",
@@ -19158,6 +17884,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"fr",
@@ -19284,6 +18017,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ja",
@@ -19473,6 +18213,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ru",
@@ -19511,9 +18258,16 @@
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"sw",
-    "task":"classification",
+    "task":"classification",
+    "metric":"accuracy",
+    "score":0.9
+  },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"sw",
+    "task":"mgsm",
     "metric":"accuracy",
-    "score":0.9
+    "score":0.7
   },
   {
     "model":"mistralai\/mistral-saba",
@@ -19557,6 +18311,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"te",
@@ -19676,6 +18437,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"zh",
@@ -19760,6 +18528,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"bn",
@@ -19802,6 +18577,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"de",
@@ -19844,6 +18626,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"en",
@@ -19886,6 +18675,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"es",
@@ -19928,6 +18724,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"fr",
@@ -20054,6 +18857,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ja",
@@ -20171,14 +18981,14 @@
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"bleu",
-    "score":0.3668684839
+    "score":0.3647734864
   },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"pa",
     "task":"translation_from",
     "metric":"chrf",
-    "score":0.5826633325
+    "score":0.5784168493
   },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
@@ -20243,6 +19053,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ru",
@@ -20285,6 +19102,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"sw",
@@ -20327,6 +19151,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"te",
@@ -20446,6 +19277,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"zh",
@@ -20530,6 +19368,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"bn",
@@ -20572,6 +19417,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"de",
@@ -20614,6 +19466,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"en",
@@ -20656,6 +19515,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"es",
@@ -20698,6 +19564,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"fr",
@@ -20824,6 +19697,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ja",
@@ -21013,6 +19893,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ru",
@@ -21055,6 +19942,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"sw",
@@ -21097,6 +19991,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"te",
@@ -21216,6 +20117,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"zh",
@@ -21300,6 +20208,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"bn",
@@ -21342,6 +20257,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"de",
@@ -21384,6 +20306,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"en",
@@ -21426,6 +20355,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"es",
@@ -21468,6 +20404,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"fr",
@@ -21536,14 +20479,14 @@
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"bleu",
-    "score":0.3803524524
+    "score":0.3858833658
   },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"hi",
     "task":"translation_to",
     "metric":"chrf",
-    "score":0.5965079449
+    "score":0.5991711103
   },
   {
     "model":"openai\/gpt-4.1-nano",
@@ -21594,6 +20537,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ja",
@@ -21783,6 +20733,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ru",
@@ -21825,6 +20782,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"sw",
@@ -21867,6 +20831,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"te",
@@ -21986,6 +20957,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"zh",
@@ -22070,6 +21048,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"bn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"bn",
@@ -22112,6 +21097,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"de",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"de",
@@ -22154,6 +21146,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"en",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"en",
@@ -22196,6 +21195,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"es",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"es",
@@ -22238,6 +21244,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"fr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"fr",
@@ -22364,6 +21377,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ja",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ja",
@@ -22553,6 +21573,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ru",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ru",
@@ -22595,6 +21622,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"sw",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"sw",
@@ -22637,6 +21671,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"te",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"te",
@@ -22756,6 +21797,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"zh",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"zh",