README.md

metadata

tags:
  - hetu
  - mathreasoning
  - cotgen

HETU-MathReasoning-CotGen-Data

Training data for the HETU (Hints Enable True Understanding) paper.

• Task: math reasoning (AIME, GSM8K, MATH-500, Omni-MATH, GPQA-Diamond, MMLU)
• Method: CotGen
• Examples: 3,481 train / 35-36 held-out val
• Format: ShareGPT (system + conversations: [{from, value}]), used for LoRA SFT via LLaMA-Factory.

Single-turn ShareGPT conversations, curated DeepSeek-R1-style math reasoning distillation. Each row: the raw math/logic problem (human turn) and the full teacher chain-of-thought plus a boxed final answer (gpt turn), with loss computed over the entire gpt turn.

Files:

• train.json -- training split
• val.json -- held-out validation split

See the HETU paper for full dataset construction methodology, and the corresponding HETU-*-MathReasoning-CotGen model repos for checkpoints trained on this data.