metadata
license: mit
LM optimization: reasoning documents
Generated reasoning documents from https://github.com/LamShiuChing/LM-optimization. The datasets are recipes
(src/tasks.py generates them from a seed); each <version>.jsonl here is a 100k-document sample of one
version, one JSON object per line with task, prompt, cot (the step trace) and answer. A document
trains as prompt|cot=answer. The format is described in docs/language.md and docs/dataset.md of the repo.