AI Potluck
Model components / Training & synthetic datasets

DCLM-baseline

DataComp-LM (ML Foundations)

Curated pretraining dataset derived from Common Crawl, about 4 trillion tokens across 3 billion documents and roughly 7.2TB, produced by the DataComp-LM project. It was built with heuristic cleaning, Bloom-filter deduplication and fastText model-based quality filtering, and is the reference dataset for the DCLM benchmark.

OLMoE, SmolLM3, Marin and RWKV were built on it. Verified 2026-08-13 via the mlfoundations/dclm-baseline-1.0 dataset card on Hugging Face.

Openness

5 high confidence
5.0
license
CC-BY-4.0
gated
false
dataset_card
present
github
mlfoundations/dclm

Permissive CC-BY-4.0 license, ungated download, with open code repo and dataset card.

Adoption

4 high confidence
4.0

309,488 downloads in the trailing 30 days for mlfoundations/dclm-baseline-1.0, which bands at 100K-1M, level 4 on the dataset adoption scale.

Capability

5 high confidence
5.0

Leads the DataComp-LM leaderboard (beats C4/RefinedWeb/RedPajama/Dolma); current 2024-26 default of OLMoE, Marin, SmolLM3, and RWKV. DCLM-Baseline 7B reaches 64% 5-shot MMLU at 2.6T tokens, a 6.6-point MMLU gain over MAP-Neo, the prior open-data state of the art.

Verified 2026-08-13