| # Model V — generative process verifier / first-error localizer. | |
| # Inherits configs/base.yaml; keys here override. | |
| extends: "base.yaml" | |
| task: "v" | |
| output_dir: "runs/verifier_v" | |
| hub_model_id: null # e.g. "your-username/mathcompose-verifier-1.5b" | |
| data: | |
| train_path: "data/verifier/train.jsonl" | |
| val_path: "data/verifier/val.jsonl" | |
| # jsonl rows are {"prompt": <chat messages or str>, "completion": <critique + \boxed{k}>} | |
| train: | |
| # Verifier inputs (problem + full solution + per-step critique) are long: | |
| # push seq_len toward the 4096 ceiling. Report truncation rate at build time. | |
| max_length: 3072 | |
| per_device_train_batch_size: 1 | |
| gradient_accumulation_steps: 16 | |
| eval: | |
| processbench_splits: ["gsm8k", "math", "olympiadbench", "omnimath"] | |
| maj_k: 8 # Maj@k test-time voting for the headline number | |
| gpt4o_reference_f1: 61.9 # standard proprietary baseline to beat | |