MiniMax-M3-150k-Mixed

# m3-alldomains-verified-107k Verified distillation traces generated with [faststill](https://empero.org) **v0.0.1** — a pipeline that generates `(prompt, reasoning, output)` triplets from any OpenAI-compatible chat-completions endpoint and **deterministically verifies every row** before keeping it. A row is `verified=true` only when a machine check (executed unit tests, exact / normalized answer compare) confirmed it, so wrong labels are filtered out instead of poisoning a student model. ## Dataset summary - **Rows:** 107,060 (all verified) - **Tasks:** code, ifeval, math, math_latex, mc - **Source prompt sets:** amc_aime, aops_forum, arc, cn_k12, evalplus, gsm8k, humaneval, ifeval, kodcode, limo, math, mbpp, medmcqa, medqa, metamath, mmlu, mmlu_pro, olympiads, openbookqa, orca_math, pubmedqa, sciq, synthetic_amc, synthetic_math, theoremqa - **Teacher model(s):** MiniMax-M3 - **Generated:** 2026-06-19T06:56:36.325801+00:00 - **Generator:** faststill v0.0.1 ## Using it for SFT Each row separates the model's chain-of-thought (`reasoning`) from its final answer/function (`output`). For supervised fine-tuning, build the assistant turn from `reasoning` + `output` on rows where `verified == true`. ```python from datasets import load_dataset ds = load_dataset("json", data_files="*.jsonl", split="train") ds = ds.map(lambda r: {"messages": [ {"role": "user", "content": r["prompt"]}, {"role": "assistant", "content": r["reasoning"] + "\n\n" + r["output"]}, ]}) ``` ## Fields | field | description | |---|---| | `prompt_id` | stable id of the source prompt | | `task` | verification task: `math`, `math_latex`, `code`, `mc`, `ifeval` | | `prompt` | the input question | | `reasoning` | model chain-of-thought (for reasoning models, the `<think>` trace; tags stripped) | | `output` | the final answer / full runnable function | | `completion` | full raw model output (audit) | | `ground_truth` | the reference answer / canonical solution | | `verified` | **true only if the deterministic check passed** | | `score` | 0.0–1.0 (1.0 = fully verified) | | `reason` | verification detail (e.g. `all_tests_passed`, `extracted=42 expected=42`) | | `deterministic_verified / judge_*` | optional LLM-judge fields when `--judge` was used | | `prompt_tokens / completion_tokens / latency_s / model` | generation metadata | | `meta` | free-form source metadata | ## Statistics | task | rows | verified | pass rate | what it is | |---|---:|---:|---:|---| | `code` | 15,922 | 15,922 | 100% | Python code — verified by executing the unit tests in a sandboxed subprocess | | `ifeval` | 47 | 47 | 100% | verifiable instruction-following constraints | | `math` | 21,154 | 21,154 | 100% | grade-school math — verified by `#### N` exact-answer compare | | `math_latex` | 15,771 | 15,771 | 100% | competition math — verified by `\boxed{}` extraction + LaTeX normalization | | `mc` | 54,166 | 54,166 | 100% | multiple choice — verified by strict A/B/C/D extraction | Tokens — prompt: 30,366,412 · completion: 72,197,326 ## Verification & limitations - **Code** is executed against the dataset's unit tests in a sandboxed subprocess (resource limits, scrubbed env, process-group kill, a static reward-hack scanner). It is defense-in-depth, not a VM. - **Math** anchors to the final answer and compares exactly (integers) / tightly (floats); **math_latex** is case-sensitive and delimiter-preserving. - The verifier fails **closed**: a borderline row is dropped, not kept — so `verified=false` rows include genuinely-wrong outputs *and* outputs the checker couldn't parse. Filter to `verified == true` for training. ## Support / Donate If this tool helped you, consider supporting the project: - **BTC:** `bc1qx6zepu6sfkvshgdmc4ewu6pk6rpadvpgffpp7v` - **LTC:** `ltc1qv2mefzps2vtjcpwfx8xxdrpplrcvltswm68r7x` - **XMR:** `42Dbm5xg5Nq26fdyzfEU7KBnAJfhi7Cvz5J2ex5CzHXkfKuNEJzYCcmJ1GTbgjFZ5MBx72sdG1G9239Cd6rsZfv4QeDkYJY` ## Links - Project: [https://empero.org](https://empero.org) - Generated with [faststill](https://empero.org) v0.0.1