README.md

metadata

pretty_name: OpenThoughts3 Math
task_categories:
  - text-generation
tags:
  - math
  - reasoning
  - sft
size_categories:
  - 100K<n<1M

OpenThoughts3 Math

This dataset contains the math-only, complete-solution subset used for supervised fine-tuning in LLM-Fusion experiments. It was derived from open-thoughts/OpenThoughts3-1.2M.

Dataset summary

• 103,760 training rows
• 32,193 unique math questions
• Up to four solutions per question, selected deterministically with seed 20260910
• All rows have domain = "math" and source = "ai2-adapt-dev/openmath-2-math"
• Solutions are retained only when the assistant output contains a closing </think> tag and a \boxed{...} answer after it
• Maximum rendered sequence length: 32,768 tokens with the Qwen3-4B-Instruct-2507 tokenizer

The completeness filter favors questions whose generated solutions finish within the original generation limit, so this subset should not be treated as an unbiased sample of the source dataset's difficulty distribution.

Schema

Column	Type	Description
messages	list of {role, content}	User prompt and assistant solution
domain	string	Always math
source	string	Original source identifier
dataset	string	OpenThoughts3-1.2M
split	string	Always train
difficulty	string	Optional difficulty metadata
sample_id	string	Stable sample identifier

Construction statistics

The builder scanned 1,200,000 source rows. It excluded 350,000 non-math rows, incomplete generations, generations without a boxed final answer, and extra solutions above the four-per-question cap.