What it is
234K scientific QA pairs built from paper-level reasoning skeletons rather than raw text, targeting hard, self-contained multi-step reasoning for SFT. 234K synthesized QA pairs from ~370K filtered papers across 10 disciplines, built via claim-centered reasoning skeletons that encode claims, evidence, quantities, assumptions, and limits.
Key results
- 234K English QA pairs in JSONL
- ~370K sampled papers from a 3.36M-paper source base
- 10 disciplines; Apache-2.0 license
Useful for supervised fine-tuning of LLMs on hard, self-contained scientific reasoning.