42e/mathcompose-verifier-7b

mathcompose — Model V (process verifier) SFT data

First-error-localization critiques for training a small generative math verifier. Each row: a problem + a step-indexed candidate solution (prompt) and a paragraph-by-paragraph critique ending in \boxed{{k}} (completion), where k is the 0-based index of the first erroneous step (-1 = all correct).

• Labels are ground truth from PRM800K (OpenAI, MIT).
• Genuine-detection recipe: claude-opus-4-8 critiques each solution blind (never shown the answer); a critique is kept only if its predicted index matches the PRM800K gold (rejection sampling, ~77% keep-rate). So the critiques contain real error-finding reasoning, not rationalization.
• Deduped against ProcessBench + MATH-500 (no eval contamination).
• Naturally ~50/50 erroneous/all-correct (no rebalancing needed).

Built by the mathcompose harness. Eval target: ProcessBench first-error F1.