Skip to content

Dataset · HuggingFaceTB

Math educational corpus from CommonCrawl

HuggingFaceTB/finemath

34-54B tokens of math educational texts filtered from CommonCrawl for model training.

What's inside

Contains math educational content curated from CommonCrawl. A classifier was trained on Llama-3.1-70B-Instruct annotations for filtering, keeping clear explanations and step-by-step problem solutions. The volume is 34B tokens in FineMath-3+ and 54B tokens together with InfiMM-WebMath-3+. The language is not specified in the card; metadata classifies the data as tabular and text.

What you can do with it

  • Continued pretraining of a mathematical LLM
  • SFT on step-by-step mathematical explanations
  • Training a mathematical content classifier
Size
10M–100M rows
Downloads this month
70.1K
Likes
400

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("HuggingFaceTB/finemath")