Update to new Mars-Bench dataset structure with Author field

app.py

@@ -49,7 +49,7 @@ def build_tab(df, name):
     datasets = sorted(df["Dataset"].unique().tolist())
     models = sorted(df["Model"].unique().tolist())
     organizations = sorted(df["Organization"].unique().tolist())
-    metric_cols = [col for col in df.columns if col not in ["Model", "Organization", "Dataset", "First Author", "Author Link"]]
+    metric_cols = [col for col in df.columns if col not in ["Model", "Organization", "Dataset", "Author", "Author Link"]]
 
     # Create pivoted dataframe
     pivoted_data = []
@@ -57,17 +57,17 @@ def build_tab(df, name):
         for org in organizations:
             model_org_data = df[(df["Model"] == model) & (df["Organization"] == org)]
             if not model_org_data.empty:
-                # Get First Author and Author Link from the first entry (they should be the same for model+org)
-                first_author = model_org_data["First Author"].values[0]
+                # Get Author and Author Link from the first entry (they should be the same for model+org)
+                author = model_org_data["Author"].values[0]
                 author_link = model_org_data["Author Link"].values[0]
 
                 # Format as markdown link if author_link exists
                 if pd.notna(author_link) and author_link.strip():
-                    author_display = f"[{first_author}]({author_link})"
+                    author_display = f"[{author}]({author_link})"
                 else:
-                    author_display = first_author
+                    author_display = author
 
-                row = {"Model": model, "Organization": org, "First Author": author_display}
+                row = {"Model": model, "Organization": org, "Author": author_display}
                 for dataset in datasets:
                     dataset_data = model_org_data[model_org_data["Dataset"] == dataset]
                     if not dataset_data.empty:
@@ -88,7 +88,7 @@ def build_tab(df, name):
         for metric in metric_cols:
             metric_combo_cols.append(f"{dataset} ({metric})")
 
-    all_cols = ["Model", "Organization", "First Author"] + metric_combo_cols
+    all_cols = ["Model", "Organization", "Author"] + metric_combo_cols
 
     with gr.TabItem(name, elem_id="llm-benchmark-tab-table"):
         with gr.Row():
@@ -100,7 +100,7 @@ def build_tab(df, name):
                 )
 
                 # Column selector for base columns only (not dataset+metric combos)
-                base_cols = ["Model", "Organization", "First Author"]
+                base_cols = ["Model", "Organization", "Author"]
                 col_selector = gr.CheckboxGroup(
                     choices=base_cols,
                     value=base_cols,
@@ -108,10 +108,10 @@ def build_tab(df, name):
                     elem_classes="column-select"
                 )
 
-                # Set datatype to 'markdown' for First Author column to enable clickable links
+                # Set datatype to 'markdown' for Author column to enable clickable links
                 datatypes = []
                 for col in pivoted_df.columns:
-                    if col == "First Author":
+                    if col == "Author":
                         datatypes.append("markdown")
                     else:
                         datatypes.append("str")
@@ -246,7 +246,7 @@ custom_css = """
     min-width: 120px;
 }
 
-/* Style links in First Author column */
+/* Style links in Author column */
 #leaderboard-table a {
     color: #0066cc;
     text-decoration: none;
@@ -354,10 +354,10 @@ with demo:
                     )
 
                 with gr.Column():
-                    submit_first_author = gr.Textbox(
-                        label="First Author",
+                    submit_author = gr.Textbox(
+                        label="Author",
                         placeholder="e.g., K. He",
-                        info="Name of the first author"
+                        info="Name of the author"
                     )
                     submit_author_link = gr.Textbox(
                         label="Author Page Link (Optional)",
@@ -404,9 +404,9 @@ with demo:
             We'll review your submission and add it to the leaderboard if approved.
             """)
 
-            def generate_submission_text(task, model, org, dataset, first_author, author_link, metrics, paper, code, email, notes):
-                if not all([task, model, org, dataset, first_author, metrics, email]):
-                    return "❌ Error: Please fill in all required fields (Task Type, Model Name, Organization, Dataset, First Author, Metrics, Contact Email)"
+            def generate_submission_text(task, model, org, dataset, author, author_link, metrics, paper, code, email, notes):
+                if not all([task, model, org, dataset, author, metrics, email]):
+                    return "❌ Error: Please fill in all required fields (Task Type, Model Name, Organization, Dataset, Author, Metrics, Contact Email)"
 
                 submission_text = f"""## New Model Submission
 
@@ -414,7 +414,7 @@ with demo:
 **Model Name:** {model}
 **Organization:** {org}
 **Dataset:** {dataset}
-**First Author:** {first_author}
+**Author:** {author}
 **Author Page Link:** {author_link if author_link else "N/A"}
 
 ### Metrics
@@ -439,7 +439,7 @@ with demo:
 
             generate_btn.click(
                 generate_submission_text,
-                inputs=[submit_task, submit_model, submit_org, submit_dataset, submit_first_author,
+                inputs=[submit_task, submit_model, submit_org, submit_dataset, submit_author,
                        submit_author_link, submit_metrics, submit_paper, submit_code, submit_email, submit_notes],
                 outputs=submission_output
             )

data/classification.csv

@@ -1,16 +1,10 @@
-Model,Organization,Dataset,Accuracy,F1-Score,First Author,Author Link
-ResNet-50,Microsoft,DoMars16,92.5,91.8,K. He,https://kaiminghe.github.io/
-ViT-Base,Google,DoMars16,94.2,93.5,A. Dosovitskiy,https://scholar.google.com/citations?user=UfvnhKcAAAAJ
-Swin-T,Microsoft,DoMars16,95.8,94.9,Z. Liu,https://scholar.google.com/citations?user=nHh9PSsAAAAJ
-InceptionV3,Google,DoMars16,93.1,92.4,C. Szegedy,https://scholar.google.com/citations?user=3QeF7mAAAAAJ
-SqueezeNet,DeepMind,DoMars16,89.7,88.6,F. Iandola,https://scholar.google.com/citations?user=LvWAIwUAAAAJ
-ResNet-50,Microsoft,Atmospheric Dust,88.3,87.5,K. He,https://kaiminghe.github.io/
-ViT-Base,Google,Atmospheric Dust,90.1,89.2,A. Dosovitskiy,https://scholar.google.com/citations?user=UfvnhKcAAAAJ
-Swin-T,Microsoft,Atmospheric Dust,91.5,90.7,Z. Liu,https://scholar.google.com/citations?user=nHh9PSsAAAAJ
-InceptionV3,Google,Atmospheric Dust,89.8,88.9,C. Szegedy,https://scholar.google.com/citations?user=3QeF7mAAAAAJ
-SqueezeNet,DeepMind,Atmospheric Dust,87.2,86.3,F. Iandola,https://scholar.google.com/citations?user=LvWAIwUAAAAJ
-ResNet-50,Microsoft,Martian Frost,85.6,84.8,K. He,https://kaiminghe.github.io/
-ViT-Base,Google,Martian Frost,87.9,86.9,A. Dosovitskiy,https://scholar.google.com/citations?user=UfvnhKcAAAAJ
-Swin-T,Microsoft,Martian Frost,88.4,87.5,Z. Liu,https://scholar.google.com/citations?user=nHh9PSsAAAAJ
-InceptionV3,Google,Martian Frost,86.7,85.8,C. Szegedy,https://scholar.google.com/citations?user=3QeF7mAAAAAJ
-SqueezeNet,DeepMind,Martian Frost,84.3,83.4,F. Iandola,https://scholar.google.com/citations?user=LvWAIwUAAAAJ
+Model,Organization,Dataset,Accuracy,F1-Score,Author,Author Link
+Under Construction,Under Construction,atmospheric_dust_cls_edr,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,atmospheric_dust_cls_rdr,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,change_cls_ctx,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,change_cls_hirise,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,domars16k,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,frost_cls,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,landmark_cls,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,surface_cls,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,surface_multi_label_cls,Under Construction,Under Construction,Under Construction,

data/detection.csv

@@ -1,16 +1,4 @@
-Model,Organization,Dataset,mAP,IoU,First Author,Author Link
-Faster R-CNN,Meta,Mars Crater,78.5,0.72,S. Ren,https://scholar.google.com/citations?user=L_hq4a0AAAAJ
-YOLOv5,Ultralytics,Mars Crater,80.2,0.74,G. Jocher,https://github.com/glenn-jocher
-DETR,Meta,Mars Crater,82.1,0.76,N. Carion,https://scholar.google.com/citations?user=p1wMCUUAAAAJ
-RetinaNet,Meta,Mars Crater,79.3,0.73,T. Lin,https://scholar.google.com/citations?user=7zQoaFkAAAAJ
-SSD,Google,Mars Crater,77.8,0.71,W. Liu,https://scholar.google.com/citations?user=PQ7QBjYAAAAJ
-Faster R-CNN,Meta,Rover Component,75.6,0.69,S. Ren,https://scholar.google.com/citations?user=L_hq4a0AAAAJ
-YOLOv5,Ultralytics,Rover Component,77.3,0.71,G. Jocher,https://github.com/glenn-jocher
-DETR,Meta,Rover Component,78.9,0.73,N. Carion,https://scholar.google.com/citations?user=p1wMCUUAAAAJ
-RetinaNet,Meta,Rover Component,76.7,0.70,T. Lin,https://scholar.google.com/citations?user=7zQoaFkAAAAJ
-SSD,Google,Rover Component,75.1,0.68,W. Liu,https://scholar.google.com/citations?user=PQ7QBjYAAAAJ
-Faster R-CNN,Meta,Geological Feature,73.4,0.67,S. Ren,https://scholar.google.com/citations?user=L_hq4a0AAAAJ
-YOLOv5,Ultralytics,Geological Feature,75.1,0.69,G. Jocher,https://github.com/glenn-jocher
-DETR,Meta,Geological Feature,76.7,0.71,N. Carion,https://scholar.google.com/citations?user=p1wMCUUAAAAJ
-RetinaNet,Meta,Geological Feature,74.5,0.68,T. Lin,https://scholar.google.com/citations?user=7zQoaFkAAAAJ
-SSD,Google,Geological Feature,73.0,0.67,W. Liu,https://scholar.google.com/citations?user=PQ7QBjYAAAAJ
+Model,Organization,Dataset,mAP,IoU,Author,Author Link
+Under Construction,Under Construction,boulder_det,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,conequest_det,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,dust_devil_det,Under Construction,Under Construction,Under Construction,

data/segmentation.csv

@@ -1,16 +1,9 @@
-Model,Organization,Dataset,Dice Score,IoU,First Author,Author Link
-U-Net,OpenAI,Mars Terrain,0.85,0.74,O. Ronneberger,https://scholar.google.com/citations?user=MXXkjXkAAAAJ
-DeepLabV3+,Google,Mars Terrain,0.87,0.76,L. Chen,https://scholar.google.com/citations?user=WicCTRsAAAAJ
-Mask R-CNN,Meta,Mars Terrain,0.88,0.78,K. He,https://kaiminghe.github.io/
-SegFormer,NVIDIA,Mars Terrain,0.86,0.75,E. Xie,https://scholar.google.com/citations?user=txgkZ0kAAAAJ
-HRNet,Microsoft,Mars Terrain,0.84,0.73,J. Wang,https://scholar.google.com/citations?user=aYKQn88AAAAJ
-U-Net,OpenAI,Dust Storm,0.82,0.70,O. Ronneberger,https://scholar.google.com/citations?user=MXXkjXkAAAAJ
-DeepLabV3+,Google,Dust Storm,0.84,0.72,L. Chen,https://scholar.google.com/citations?user=WicCTRsAAAAJ
-Mask R-CNN,Meta,Dust Storm,0.85,0.74,K. He,https://kaiminghe.github.io/
-SegFormer,NVIDIA,Dust Storm,0.83,0.71,E. Xie,https://scholar.google.com/citations?user=txgkZ0kAAAAJ
-HRNet,Microsoft,Dust Storm,0.82,0.70,J. Wang,https://scholar.google.com/citations?user=aYKQn88AAAAJ
-U-Net,OpenAI,Geological Feature,0.81,0.68,O. Ronneberger,https://scholar.google.com/citations?user=MXXkjXkAAAAJ
-DeepLabV3+,Google,Geological Feature,0.83,0.70,L. Chen,https://scholar.google.com/citations?user=WicCTRsAAAAJ
-Mask R-CNN,Meta,Geological Feature,0.84,0.72,K. He,https://kaiminghe.github.io/
-SegFormer,NVIDIA,Geological Feature,0.82,0.69,E. Xie,https://scholar.google.com/citations?user=txgkZ0kAAAAJ
-HRNet,Microsoft,Geological Feature,0.81,0.68,J. Wang,https://scholar.google.com/citations?user=aYKQn88AAAAJ
+Model,Organization,Dataset,Dice Score,IoU,Author,Author Link
+Under Construction,Under Construction,boulder_seg,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,conequest_seg,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,crater_binary_seg,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,crater_multi_seg,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,mars_seg_mer,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,mars_seg_msl,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,mmls,Under Construction,Under Construction,Under Construction,
+Under Construction,Under Construction,s5mars,Under Construction,Under Construction,Under Construction,