A large reasoning trace is not necessarily a transparent record of how a model reached its answer. This corpus is most useful as a study of generated training material: it pairs long-form reasoning with final responses across diverse prompts, while explicitly leaving correctness and the relationship to internal computation unverified.
What Sets It Apart
- The 250 prompts are reused from an existing high-reasoning prompt set, so the responses can be compared with other models’ answers to the same inputs; this is a response corpus, not a newly designed benchmark.
- It contains about 1.67 million output tokens, with a mean of 6,691 per prompt and a maximum of 43,473. That scale makes it useful for analyzing long outputs, but length alone does not establish reasoning quality.
- Each assistant record separates
reasoningfromcontent, alongside prompt and usage metadata. This supports studies of trace formatting, response profiles, and generation costs. - Locally assigned labels cover coding, computer science, statistics, finance, business, science, and humanities; 105 prompts remain uncategorized, so category-level comparisons are uneven.
Who It’s For and Trade-offs
Great fit for researchers and practitioners examining synthetic reasoning data, distillation, or model response styles. Look elsewhere for a correctness-verified evaluation set or a balanced benchmark: outputs are unchecked, reasoning text should not be treated as a faithful account of internal computation, and upstream data and model terms require review before redistribution or commercial use.