A curated, multi-domain corpus of condensed reasoning traces in
thinking-cap style. Each row is a complete training unit —
prompt →
condensed reasoning → numbered answer — built to
post-train reasoning models with SFT → RAFT → DPO in mind.
Goal: teach a model how to think efficiently — analyze, decompose,
compute, verify, conclude — in few tokens, while covering enough breadth
(math, science, tool calls, instruction following… See the full description on the dataset page:
https://huggingface.co/datasets/osk-arr00/thinkingcap-reasoning-traces.
