Model components / Training & synthetic datasets

FineMath

FineMath is Hugging Face's classifier-filtered mathematical web corpus extracted from Common Crawl (tens of billions of tokens at the 3+/4+ quality tiers), released under ODC-BY. It is a documented pretraining component of SmolLM3, Apertus, Marin, and OLMo 3's Dolma 3 pool.

Openness

5 high confidence

5.0

license: ODC-BY
gated: false
dataset_card: present

ODC-BY, ungated, full dataset card.

https://huggingface.co/datasets/HuggingFaceTB/finemath recorded 2026-07-04
ODC-BY license, ungated, dataset card

Adoption

3 high confidence

3.0

24,077 monthly downloads on Hugging Face, graded on the training-corpus bands.

https://huggingface.co/api/datasets/HuggingFaceTB/finemath recorded 2026-07-04
downloads field = 24,077 (30-day)

Capability

5 high confidence

5.0

Leads controlled math ablations (FineMath-4+ ~2x GSM8K and ~6x MATH over InfiMM-WebMath, beats OpenWebMath); current 2025 default in SmolLM3, Apertus, Marin, and OLMo 3.

https://huggingface.co/datasets/HuggingFaceTB/finemath recorded 2026-07-04
FineMath-4+ achieves ~2x GSM8K and ~6x MATH over InfiMM-WebMath and beats OpenWebMath
https://arxiv.org/html/2502.02737v1 recorded 2026-07-04
SmolLM2 paper details the FineMath ablation vs OpenWebMath/InfiMM-WebMath

Unchanged since 2026-07-04 (last edited, not re-checked)