panzs19/repro-organic-data-72B-fasttext-split

RePro Organic Data 72B FastText Split

This directory contains a local split of cx-cmu/repro-organic-data-72B using the RePro/DCLM fastText quality classifier.

Split Definition

The RePro paper uses DCLM-fastText with threshold 0.018112:

• high_quality_subset/: fasttext_oh_eli5_vs_rw_v2_prob >= 0.018112
• recycled_data_pool/: fasttext_oh_eli5_vs_rw_v2_prob < 0.018112

Here, recycled_data_pool means the low-quality organic pool to be recycled or used for graph construction. It is not the already rewritten/rephrased output data.

The original downloaded shards did not contain this quality score, so the score was recomputed with RePro's fasttext_oh_eli5.bin model. Each output record preserves the original JSON fields and adds fasttext_oh_eli5_vs_rw_v2_prob.

Directory Layout

repro-organic-data-72B-fasttext-split/
  high_quality_subset/
    shard_00000000_processed.jsonl.zstd
    shard_00000000_processed.jsonl.zstd.done
    ...
  recycled_data_pool/
    shard_00000000_processed.jsonl.zstd
    shard_00000000_processed.jsonl.zstd.done
    ...
  split_config.json
  split_summary.json
  split_stats.jsonl

Only files matching shard_*.jsonl.zstd are data files. The .done files are completion markers.

Summary

• Input shards: 600
• Total documents: 53,435,703
• High-quality documents: 5,584,841
• Recycled-pool documents: 47,850,862
• High-quality document ratio: 10.45%
• Estimated high-quality word ratio by previous_word_count: 17.67%

See split_summary.json and split_stats.jsonl for machine-readable statistics.

Record Format

Each line is one JSON object compressed with zstd. The schema is:

• text: page text.
• url: source URL.
• metadata: WARC / HTTP metadata dictionary.
• warcinfo: Common Crawl WARC info string.
• previous_word_count: original word-count metadata.
• language_id_whole_page_fasttext: language ID probabilities, usually like {"en": 0.98}.
• bff_contained_ngram_count_before_dedupe: deduplication-related count from the original DCLM-RefinedWeb processing.
• fasttext_oh_eli5_vs_rw_v2_prob: recomputed RePro/DCLM fastText quality score used for this split.

Because the fastText probability is reconstructed from floating point output, it may be slightly below 0 or slightly above 1 for extreme examples. The split uses the raw score and the threshold above, without clipping.

Loading With HuggingFace Datasets

Use streaming mode for normal work. Loading the full split into memory/cache is large.

from datasets import load_dataset

base = "/root/ceph/research/GraphSynData/data/repro-organic-data-72B-fasttext-split"

ds_low = load_dataset(
    "json",
    data_files=f"{base}/recycled_data_pool/shard_*.jsonl.zstd",
    split="train",
    streaming=True,
)

example = next(iter(ds_low))
print(example.keys())
print(example["text"][:200])
print(example["fasttext_oh_eli5_vs_rw_v2_prob"])

To load both subsets:

from datasets import load_dataset

base = "/root/ceph/research/GraphSynData/data/repro-organic-data-72B-fasttext-split"

ds = load_dataset(
    "json",
    data_files={
        "high_quality": f"{base}/high_quality_subset/shard_*.jsonl.zstd",
        "recycled": f"{base}/recycled_data_pool/shard_*.jsonl.zstd",
    },
    streaming=True,
)

high_example = next(iter(ds["high_quality"]))
low_example = next(iter(ds["recycled"]))

Manual Streaming Loader

This avoids HuggingFace cache creation and is useful for lightweight preprocessing jobs.

from pathlib import Path
import io

import orjson
import zstandard as zstd

def iter_jsonl_zstd(path):
    with Path(path).open("rb") as fh:
        reader = zstd.ZstdDecompressor().stream_reader(fh)
        text_reader = io.TextIOWrapper(reader, encoding="utf-8")
        for line in text_reader:
            if line.strip():
                yield orjson.loads(line)

path = "/root/ceph/research/GraphSynData/data/repro-organic-data-72B-fasttext-split/recycled_data_pool/shard_00000000_processed.jsonl.zstd"
for obj in iter_jsonl_zstd(path):
    print(obj["url"], obj["fasttext_oh_eli5_vs_rw_v2_prob"])
    break

Reproduction

The split was produced by:

python /root/ceph/research/GraphSynData/scripts/split_repro_organic_by_fasttext.py --workers 128 --overwrite

The script uses:

• model: /root/ceph/research/GraphSynData/related_work/repo/RePro/pretrain/baselines/mappers/enrichers/quality_prediction_enrichment_models/fasttext_oh_eli5.bin
• threshold: 0.018112
• score key: fasttext_oh_eli5_vs_rw_v2_prob