mimir-perplexity / samples_quartiles.py
versae's picture
Mdels and code
dcc5cd1
Raw
History Blame Contribute Delete
7.12 kB
import argparse
import os
from collections import defaultdict
from io import StringIO
import pandas as pd
from tqdm import tqdm
from perplexity import get_model_for
from subsampler import PerplexitySubsampler
def process_files(
directory,
reject_level,
model_override,
output_file,
group_by_prefix_lang,
prefix_lang_mapping=None,
ratio=None,
ratio_per_lang=None,
pa=None,
pb=None,
include=None,
):
if ratio or ratio_per_lang:
rows = ["doc_type,model,language,reject,bad,medium,good,norm,mean,std"]
else:
rows = ["doc_type,model,language,reject,bad,medium,good"]
files = os.listdir(directory)
grouped_files = defaultdict(list)
if prefix_lang_mapping is None:
prefix_lang_mapping = {}
# Group files by prefix and language if the option is enabled
description = "Processing files"
if group_by_prefix_lang:
description = "Processing files in groups"
for file in files:
parts = file.split('_')
prefix = parts[0]
if include and prefix not in include:
continue
lang = parts[-1].split(".")[0][:2]
group_key = prefix_lang_mapping.get(f"{prefix}_{lang}", f"{prefix}_{lang}")
grouped_files[group_key].append(file)
file_groups = grouped_files.values()
else:
file_groups = []
for file in files: # Each file is its own group
if include and not any(file.startswith(prefix) for prefix in include):
continue
file_groups.append([file])
if output_file:
progress = tqdm(file_groups, desc=description)
else:
progress = file_groups
print(rows[0])
# Process each group of files
for group in progress:
combined_perplexities = pd.DataFrame()
doc_type, lang = None, None
for file in group:
if not doc_type or not lang: # Set doc_type and lang based on the first file
parts = file.split('_')
doc_type = file.split('_')[0]
lang = parts[-1].split(".")[0][:2]
doc_type, lang = prefix_lang_mapping.get(f"{doc_type}_{lang}", f"{doc_type}_{lang}").rsplit("_", 1)
perp = pd.read_json(os.path.join(directory, file), lines=True)
perplexities = pd.read_json(StringIO(perp["perplexities"].to_json(lines=True, orient="records")), lines=True)
combined_perplexities = pd.concat([combined_perplexities, perplexities], ignore_index=True)
if model_override:
model = model_override
else:
model, _ = get_model_for(doc_type)
model_with_suffix = f"{model}_pp"
# Calculate quantiles for the combined perplexities of the group
reject = round(combined_perplexities[model_with_suffix].quantile(q=reject_level), 2)
bad = round(combined_perplexities[model_with_suffix].quantile(q=0.75), 2)
medium = round(combined_perplexities[model_with_suffix].quantile(q=0.50), 2)
good = round(combined_perplexities[model_with_suffix].quantile(q=0.25), 2)
if ratio:
subsampler = PerplexitySubsampler(combined_perplexities[model_with_suffix].values)
subsampler.set(ratio=ratio, pa=pa, pb=pb)
norm, mean, std = subsampler.norm, subsampler.mean, subsampler.sdev
sampling_stats = f",{norm},{mean},{std}"
elif ratio_per_lang:
subsampler = PerplexitySubsampler(combined_perplexities[model_with_suffix].values)
subsampler.set(ratio=ratio_per_lang.get(lang, ratio or 1.0), pa=pa, pb=pb)
norm, mean, std = subsampler.norm, subsampler.mean, subsampler.sdev
sampling_stats = f",{norm},{mean},{std}"
else:
sampling_stats = ""
row = f"{doc_type},{model},{lang},{reject},{bad},{medium},{good}{sampling_stats}"
if output_file:
rows.append(row)
else:
print(row)
if output_file:
with open(output_file, "w") as f:
for row in rows:
f.write(f"{row}\n")
def main():
""""
Each doc_type prefix needs to have an "no" lang, even of there's no real data.
These rows are crucial for the rest of the process.
"""
parser = argparse.ArgumentParser(description="Process files and compute perplexity metrics.")
parser.add_argument('directory', type=str, help='Directory containing the files to process')
parser.add_argument('--reject_level', type=float, default=0.95, help='Rejection quantile level (default: 0.95)')
parser.add_argument('--model_override', type=str, help='Override the model used')
parser.add_argument('--output_file', type=str, help='Output file in CSV format. If not given, prints to standard output.')
parser.add_argument('--group_by_prefix_lang', action='store_true', help='Group and calculate quantiles for files with the same prefix and language')
parser.add_argument('--overwrite_prefix_lang', type=str, help='Overwrite the assignment of languages to doc_type prefixes, e.g., "starcoder_en:starcoder_code,hplt_en:hplt_no"')
parser.add_argument('--sampling_ratio', type=float, help='Ratio of documents to keep for sampling. If passed, it generate distribution statistics (norm, mean, std) needed for sampling')
parser.add_argument('--sampling_ratio_per_lang', type=str, help='Ratio of documents per lang, e.g., "en:0.25,sv:0.34"')
parser.add_argument('--sampling_q1_prob', type=float, default=0.20, help='Probabilty for keeping documents in the Q1 range')
parser.add_argument('--sampling_q3_prob', type=float, default=0.05, help='Probabilty for keeping documents in the Q3 range')
parser.add_argument('--include', type=str, help='Comma separeted list of doc type prefixes to include')
args = parser.parse_args()
if args.sampling_ratio_per_lang:
# Turns "en: 0.25, sv : 0.34" into {'en': 0.25, 'sv': 0.34}
ratio_per_lang = dict(
(k.strip(), float(v.strip()))
for k, v in (item.split(":")
for item in args.sampling_ratio_per_lang.split(",")
)
)
else:
ratio_per_lang = None
if args.overwrite_prefix_lang:
# Turns "starcoder_en:starcoder_code,hplt_en:hplt_no" into {'starcoder_en': 'starcoder_code', 'hplt_en': 'hplt_no'}
prefix_lang_mapping = dict(
(k.strip(), v.strip())
for k, v in (item.split(":")
for item in args.overwrite_prefix_lang.split(",")
)
)
else:
prefix_lang_mapping = {}
process_files(
args.directory,
args.reject_level,
args.model_override,
args.output_file,
group_by_prefix_lang=args.group_by_prefix_lang,
prefix_lang_mapping=prefix_lang_mapping,
pa=args.sampling_q1_prob,
pb=args.sampling_q3_prob,
ratio=args.sampling_ratio,
ratio_per_lang=ratio_per_lang,
include=args.include.split(",") if args.include else None
)
if __name__ == "__main__":
main()