Datasets:

---

allenai

/

dolma3_mix-6T

like 24

Follow

Ai2 5.78k

I noticed that this dataset has a lot of duplicates (up to 6-7x repetitions). E.g. allenai/dolma3_mix-6T/data/common_crawl-crime_and_law-0019/shard_00000079.jsonl.zst is literally repeated 6 times so 83% of the file are duplicates. The same is true for many other files. I gave it to claude and it estimated the following:

  Method: Randomly sampled 200 files (out of 63,911 total) from allenai/dolma3_mix-6T, downloaded each, counted total vs unique IDs per file.

  Results:
  - 45% of files (90/200) contain duplicates
  - 2.80x overall inflation — only 35.7% of rows are unique (4.28M unique out of 11.99M total rows sampled)
  - Worst offenders: common_crawl-*-0019 sources (5-7x), common_crawl-*-0018 (2-4x), stack_edu-* (2-3x)
  - olmocr_science_pdfs-* and older common_crawl epochs (-0016, -0017) are mostly clean

  Using all 63,911 shards with dedup would give ~480B unique tokens.

  The allenai/dolma3_mix-6T-1025-7B variant has the same problem, possibly worse (sources that were clean in 6T show dupes in the 7B variant).

  Verification: Independently confirmed with raw shell tools (zstd -d + jq + sort -u, no Python) that the duplicates exist in the upstream JSONL.zst files themselves — consecutive identical rows baked into the compressed files.

Run the following script to reproduce:

"""
    uv run python preprocessing/report_upstream_dupes.py \
        --dataset allenai/dolma3_mix-6T \
        --file data/common_crawl-crime_and_law-0019/shard_00000079.jsonl.zst
"""

import argparse
import io
import json
from collections import Counter

import zstandard
from huggingface_hub import hf_hub_download


def main() -> None:
    p = argparse.ArgumentParser(description="Report upstream duplicates in a HuggingFace dataset shard.")
    p.add_argument("--dataset", required=True, help="HuggingFace dataset ID")
    p.add_argument("--file", required=True, help="Path to file within the dataset repo")
    p.add_argument("--id-column", default="id", help="ID column name (default: id)")
    args = p.parse_args()

    print(f"Downloading {args.dataset}/{args.file} ...")
    local_path = hf_hub_download(args.dataset, args.file, repo_type="dataset")

    print(f"Reading {local_path} ...")
    dctx = zstandard.ZstdDecompressor()
    rows_by_id: dict[str, list[int]] = {}
    n_rows = 0
    with open(local_path, "rb") as fh:
        with dctx.stream_reader(fh) as reader:
            for i, line in enumerate(io.TextIOWrapper(reader, encoding="utf-8")):
                record = json.loads(line)
                rid = str(record[args.id_column])
                rows_by_id.setdefault(rid, []).append(i)
                n_rows += 1

    n_unique = len(rows_by_id)
    occurrence_counts = Counter(len(lines) for lines in rows_by_id.values())

    print(f"\nFile: {args.file}")
    print(f"Total rows:  {n_rows:,}")
    print(f"Unique IDs:  {n_unique:,}")
    print(f"Duplicates:  {n_rows - n_unique:,} ({100 * (1 - n_unique / n_rows):.1f}% of rows)")
    print(f"\nOccurrence distribution:")
    for count, n_ids in sorted(occurrence_counts.items()):
        label = "unique" if count == 1 else "duplicated"
        print(f"  {count}x: {n_ids:,} IDs ({label})")

    # Check whether duplicates are consecutive and identical
    duped_ids = [(rid, lines) for rid, lines in rows_by_id.items() if len(lines) > 1]
    if duped_ids:
        sample_id, sample_lines = duped_ids[0]
        consecutive = all(sample_lines[i + 1] - sample_lines[i] == 1 for i in range(len(sample_lines) - 1))
        print(f"\nSample duplicated ID: {sample_id}")
        print(f"  Appears at lines: {sample_lines}")
        print(f"  Consecutive: {consecutive}")


if __name__ == "__main__":
    main()

which will produce

File: data/common_crawl-crime_and_law-0019/shard_00000079.jsonl.zst
Total rows:  189,780
Unique IDs:  31,630
Duplicates:  158,150 (83.3% of rows)

Occurrence distribution:
  6x: 31,630 IDs (duplicated)

Sample duplicated ID: 36c9ec53-cee5-41db-9fe3-2898c653d398
  Appears at lines: [0, 1, 2, 3, 4, 5]
  Consecutive: True