scottgeng00/dclm-8m

DCLM 8M-doc splits

Two disjoint 8,000,000-document slices of DCLM-baseline, used as a fixed, data-constrained pretraining budget (~10.1B Qwen3.5 tokens each, ~1,265 tokens/doc) in our synthetic-data / rewriting experiments:

config	documents	source rows	Qwen3.5 tokens (est.)
dclm_8m_0	8,000,000	[0, 8M)	10.10B
dclm_8m_1	8,000,000	[8M, 16M)	10.18B

Source. The DCLM portion of allenai/dolmino-mix-1124 (itself drawn from DCLM-baseline), globally shuffled; these are the first 16M rows of shuffled chunk 0, split in two. Documents are unmodified (text and id, which is the page URL, are copied verbatim).

Columns. id (URL), text, src_row (row index in the shuffled source stream).

Row order. Row j of a config is local document j of the split. Any prefix of a config is a uniform random subset (the source is shuffled), and prefixes are nested, so "the first N docs" is a well-defined smaller budget.

Reproducing our training layout. Our dataloader reads a split as 16 *.chunk.NN.jsonl files with local row j at chunk j % 16, line j // 16:

import datasets, json
ds = datasets.load_dataset("scottgeng00/dclm-8m", "dclm_8m_0", split="train")
outs = [open(f"dclm_8m_0/dclm_8m_0.chunk.{c:02d}.jsonl", "w") for c in range(16)]
for j, row in enumerate(ds):
    outs[j % 16].write(json.dumps({"id": row["id"], "text": row["text"]}) + "\n")

Held-out validation. Several hundred documents of the standard Dolmino DCLM val set share a URL or text with these splits; we evaluate on a separately built held-out set with no URL / exact / normalized-text overlap with either split.

License. DCLM-baseline is released under CC-BY-4.0 and Dolmino-mix-1124 under ODC-BY; please cite the original works:

@article{li2024datacomplm,
  title={DataComp-LM: In search of the next generation of training sets for language models},
  author={Li, Jeffrey and Fang, Alex and Smyrnis, Georgios and others},
  journal={arXiv preprint arXiv:2406.11794}, year={2024}
}
@article{olmo20242olmo2furious,
  title={2 OLMo 2 Furious}, author={OLMo, Team and others},
  journal={arXiv preprint arXiv:2501.00656}, year={2024}
}