HuggingFaceFW/finepdfs_edu_50BT-dclm_30BT-fineweb_edu_20BT

FinePDFs-Edu 50BT + DCLM 30BT + FineWeb-Edu 20BT

A ~100 billion token pretraining mixture combining three high-quality English data sources in a 50-30-20 ratio, using the educational subset of FinePDFs.

Part of the Smol-Data collection — tried and tested mixes for strong pretraining. Inspired by optimal dataset mixing.

Dataset Description

Component	Source	Tokens
FinePDFs-Edu 100BT	FinePDFs-Edu	~50B
DCLM 100BT	DCLM-Baseline 1.0	~30B
FineWeb-Edu 100BT	FineWeb-Edu	~20B

The schema is reduced to the intersection of columns across all three sources: text, id, url, and dataset.

A pre-shuffled version is available at HuggingFaceFW/finepdfs_edu_50BT-dclm_30BT-fineweb_edu_20BT-shuffled.

How It Was Created

The dataset was generated using datatrove with the smol_data.py script. Each component was subsampled from its respective 100BT subset at the target fraction (0.5, 0.3, 0.2) using a SamplerFilter with seed 42. Components were written sequentially via Slurm job dependencies to avoid concurrent commits.

Usage

from datasets import load_dataset

ds = load_dataset("HuggingFaceFW/finepdfs_edu_50BT-dclm_30BT-fineweb_edu_20BT", split="train", streaming=True)
for sample in ds:
    print(sample["text"][:200])
    break

Citation

@misc{niklaus2026smoldata,
      title={SmolData},
      author={Joel Niklaus and Hynek Kydl{\'\i}{\v{c}}ek},
      year={2026},
      publisher={Hugging Face},
      journal={Hugging Face repository},
      howpublished={\url{https://huggingface.co/collections/HuggingFaceFW/smol-data}}
}