Paired verbose and compact-register reasoning traces for 2,414 maths
problems, with per-trace follow-ability scores.
Each row holds a problem, the long chain-of-thought Qwen3-1.7B produced for it,
a compact-notation rewrite of that same reasoning, token counts for both, and
the scores used to measure how followable the compact version is to Qwen3-1.7B.
11,174,460 original think tokens → 750,087 compressed (14.9×).
