verifier-deepseek-math-7b-0622
Role: VERIFIER
Base model: deepseek-ai/deepseek-math-7b-instruct
Training data (0622 4-way split, no leakage)
| Split | Rows | Used as |
|---|---|---|
| train_rm1 | 806 | RM train (1/2) |
| train_rm2 | 805 | RM train (1/2) / Verifier test |
| train_verifier1 | 805 | Verifier train (1/2) / RM val |
| train_verifier2 | 805 | Verifier train (1/2) / RM test |
This checkpoint was trained on train_verifier1+train_verifier2 with train_rm2 as the held-out test set (and the counterpart split used as validation for best-checkpoint selection).
Train objective: SFT to emit aligned / not_aligned after the standard "You are a math solution evaluator..." system prompt.
Test-set performance (train_rm2.json, N=1610, balanced)
| Metric | Value |
|---|---|
| Accuracy | 0.8155 |
| Macro F1 | 0.8150 |
| Parse failures | 0 / 1610 |
License
Apache-2.0. Bundled training data licenses follow each source dataset (MathEDU, Stepwise Verification, MathClean, EIC).