File size: 8,165 Bytes
fe1c232 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 |
"""
- Used in RPv2 exploration, including:
- Plot certain quality signals
- Get doc/char counts (e.g. after filtering)
- Store minhashes of filtered documents for further dedup
"""
import gzip
import orjson
import numpy as np
import pandas as pd
import os
import matplotlib.pyplot as plt
from tqdm import tqdm
import copy
import multiprocessing
import random
import pathlib
import seaborn as sns
from rules.rules import gopher_rules_pass
import pyarrow
ROOT_PATH = "/home1/BharatGPT_Data/RedPajamaV2"
DATA_ROOT_PATH = "/home1/BharatGPT_Data/RedPajamaV2/data"
PLOTS_ROOT_PATH = "/home1/BharatGPT_Data/RedPajamaV2/plots"
SNAPSHOT = "2023-14"
LANGUAGE = "en"
PARTITION_KEY = "head"
SIGNALS_DIR = os.path.join(DATA_ROOT_PATH, "quality_signals", SNAPSHOT)
DUPLICATES_DIR = os.path.join(DATA_ROOT_PATH, "duplicates", SNAPSHOT)
MINHASH_DIR = os.path.join(DATA_ROOT_PATH, "minhash", SNAPSHOT)
NUM_CORES = 60
SEED = 2024
DO_PLOT = False
COUNT_ONLY = True
STORE_SIGNALS = False
OUTPUT_FILES = False
# constants for random 100
NUM_SHARDS_PROCESSED = 100
if DO_PLOT:
PLOTS_DIR = os.path.join(PLOTS_ROOT_PATH, PARTITION_KEY, f"random_{NUM_SHARDS_PROCESSED}_gopher", "quality_rep_log") # makes dir if not present
NUM_DOCS_HEAD = 2533743
NUM_DOCS_MIDDLE = 3722022
NUM_CHARS_HEAD = 12932890455
NUM_CHARS_MIDDLE = 17883781733
NUM_SHARDS_PROCESSED = 5000
if OUTPUT_FILES:
OUT_DIR = os.path.join(DATA_ROOT_PATH, f"minhash_filtered", SNAPSHOT) # makes dir if not present
random.seed(SEED)
np.random.seed(SEED)
assert sorted(os.listdir(SIGNALS_DIR)) == sorted(os.listdir(DUPLICATES_DIR))
assert (not DO_PLOT) or STORE_SIGNALS # DO_PLOT implies STORE_SIGNALS
if COUNT_ONLY:
assert DO_PLOT == False, "Plotting requires storing signal values"
all_shards_counts = {
"doc_count": 0,
"char_count": 0
}
elif STORE_SIGNALS:
all_shards_signals = {
"ccnet_perplexity": [],
# https://huggingface.co/datasets/togethercomputer/RedPajama-Data-V2
# mentions that data is raw (no deduplicated)
# but based on doc length it looks line-deduped
# so we use ccnet_length and not ccnet_original_length
"ccnet_length": [],
"rps_doc_stop_word_fraction": [],
"rps_doc_lorem_ipsum": []
}
all_shards_signals_empty = copy.deepcopy(all_shards_signals)
# non-parallelized
# for shard in tqdm(sorted(os.listdir(SIGNALS_DIR))[:10]):
# signals_path = os.path.join(SIGNALS_DIR, shard, f"{LANGUAGE}_{PARTITION}.signals.json.gz")
# duplicates_path = os.path.join(DUPLICATES_DIR, shard, f"{LANGUAGE}_{PARTITION}.duplicates.parquet")
# shard_dups = pd.read_parquet(duplicates_path)
# shard_dups_set = set(shard_dups["doc_id"].tolist())
# with gzip.open(signals_path, 'r') as signals_file:
# for line in signals_file:
# signals_dict = json.loads(line)
# if signals_dict["id"] not in shard_dups_set:
# # print(json.dumps(signals_dict, indent=4))
# for k in list(all_shards_signals.keys()):
# assert len(signals_dict["quality_signals"][k]) == 1
# assert len(signals_dict["quality_signals"][k][0]) == 3
# all_shards_signals[k].append(signals_dict["quality_signals"][k][0][-1])
partitions_dict = {
"head": ["head"],
"middle": ["middle"],
"head_middle": ["head", "middle"],
}
for partition in partitions_dict[PARTITION_KEY]:
def process_shard(shard):
# check if done already
if OUTPUT_FILES:
out_path_shard = os.path.join(OUT_DIR, shard)
out_file_path = os.path.join(out_path_shard, f"{LANGUAGE}_{partition}_filtered.minhash.parquet")
if os.path.exists(out_file_path):
raise Exception("ERROR: output file already present")
return []
minhash_path = os.path.join(MINHASH_DIR, shard, f"{LANGUAGE}_{partition}.minhash.parquet")
try:
df = pd.read_parquet(minhash_path)
except pyarrow.lib.ArrowInvalid as __e:
# occurs with empty file
print(f"ERROR with shard {shard}: empty minhash file")
return []
signals_path = os.path.join(SIGNALS_DIR, shard, f"{LANGUAGE}_{partition}.signals.json.gz")
try:
duplicates_path = os.path.join(DUPLICATES_DIR, shard, f"{LANGUAGE}_{partition}.duplicates.parquet")
shard_dups = pd.read_parquet(duplicates_path, columns=["doc_id"])
shard_dups_set = set(shard_dups["doc_id"].tolist())
except pyarrow.lib.ArrowInvalid as __e:
# occurs with empty file
shard_dups_set = set()
results = None
if COUNT_ONLY:
results = {
"doc_count": 0,
"char_count": 0
}
elif STORE_SIGNALS:
results = copy.deepcopy(all_shards_signals_empty)
elif OUTPUT_FILES:
results = [] # filtered indices
idx = -1
with gzip.open(signals_path, 'r') as signals_file:
for line in signals_file:
idx += 1
signals_dict = orjson.loads(line)
if signals_dict["id"] not in shard_dups_set and gopher_rules_pass(signals_dict):
"""
Note about exact duplicates:
https://github.com/togethercomputer/RedPajama-Data/issues/84#issuecomment-1840299911
One copy remains with this method
"""
if COUNT_ONLY:
results["doc_count"] += 1
results["char_count"] += signals_dict["quality_signals"]["ccnet_length"][0][2]
elif STORE_SIGNALS:
for k in list(results.keys()):
assert len(signals_dict["quality_signals"][k]) == 1
assert len(signals_dict["quality_signals"][k][0]) == 3
results[k].append(signals_dict["quality_signals"][k][0][2])
elif OUTPUT_FILES:
results.append(idx)
if OUTPUT_FILES:
df = df.iloc[results, :]
pathlib.Path(out_path_shard).mkdir(parents=True, exist_ok=True)
df.to_parquet(os.path.join(out_path_shard, f"{LANGUAGE}_{partition}_filtered.minhash.parquet"))
return results
with multiprocessing.Pool(NUM_CORES) as pool:
shards_list = os.listdir(SIGNALS_DIR)
all_results = list(tqdm(pool.imap(process_shard, random.sample(sorted(shards_list), k=NUM_SHARDS_PROCESSED)), total=NUM_SHARDS_PROCESSED))
for results in all_results:
if COUNT_ONLY:
for k in list(all_shards_counts.keys()):
all_shards_counts[k] += results[k]
elif STORE_SIGNALS:
for k in list(all_shards_signals.keys()):
all_shards_signals[k].extend(results[k])
# print(json.dumps(all_shards_signals))
if COUNT_ONLY:
print(all_shards_counts)
if DO_PLOT:
pathlib.Path(PLOTS_DIR).mkdir(parents=True, exist_ok=True)
for k, v in all_shards_signals.items():
# plt.hist(v, bins=100) # linear
# log with plt
# hist, bins = np.histogram(v, bins=100)
# logbins = np.logspace(np.log10(bins[0]),np.log10(bins[-1]),len(bins))
# plt.hist(v, weights=np.ones(len(v))/(NUM_DOCS_HEAD+NUM_DOCS_MIDDLE), bins=logbins) # log
# plt.xscale("log")
# unweighted
sns.histplot(x=v, bins=100, log_scale=True)
# weights: docs percentage of head_middle
# sns.histplot(x=v, weights=np.ones(len(v))*100/(NUM_DOCS_HEAD+NUM_DOCS_MIDDLE), bins=100, log_scale=True)
# weight by char percentage
# sns.histplot(x=v, weights=np.array(all_shards_signals["ccnet_length"])*100/(NUM_CHARS_HEAD+NUM_CHARS_MIDDLE), bins=100, log_scale=True)
plt.savefig(os.path.join(PLOTS_DIR, f"{k}.png"))
plt.close()
# print(len(all_shards_signals["ccnet_perplexity"]))
# print(sum(all_shards_signals["ccnet_length"]))
# print(type(all_shards_signals["ccnet_length"][0]))
|