crucible-sft-qwen3.5-4b-mini

# crucible-sft-qwen3.5-4b-mini Self-distilled SFT dataset of verified reasoning traces from **Qwen/Qwen3.5-4B**, built by the [reasoning-compression](https://github.com/marcodsn/reasoning-compression) crucible pipeline: k-sample generation on a decontaminated prompt pool, inline verification (symbolic math / sandboxed code tests), difficulty banding via solve rate, and loop-detector filtering on the chosen trace. Each row: `prompt`, `reasoning` (a verified-correct thinking trace when the domain is verifiable), `response`, `domain`, `verified`, `difficulty` (`{solve_rate, k}`), sampling params in `gen`. - Rows: **648** — domains: {"general": 219, "code": 101, "math": 191, "creative": 137} - Prompts are content-addressed (`prompt_id` = sha256 of prompt text) and decontaminated against GSM8K, MATH-500, HumanEval, MBPP, MMLU-Pro. - A stratified held-out split was carved *before* generation and is not included here. Part of the caveman reasoning-compression ablation study (see the flint datasets built from this file and the trained comparison arms). ## Build manifest ```json { "created_at": "2026-07-09T11:20:19+00:00", "git_sha": "4636916a99c191bd15795c43fbee26ac3337d1f5", "recipe": "flint-ablation-v2-mini", "recipe_def": { "target_size": 1100, "shares": { "math": 0.3, "code": 0.16, "general": 0.28, "creative": 0.26 }, "candidate_caps": { "math": 700, "code": 354, "general": 400, "creative": 350 }, "solve_band": [ 0.2, 0.8 ], "solve_band_by_domain": { "code": [ 0.2, 1.0 ] }, "k": 4, "oversample": 2.2, "lang": "en" }, "raw_rows": 1328, "selected_rows": 648, "domains": { "general": 219, "code": 101, "math": 191, "creative": 137 }, "verified": { "None": 356, "True": 292 }, "degenerate_in_raw": 272, "gen_model": "Qwen/Qwen3.5-4B" } ```