| # Model G — step-by-step solution generator. | |
| # Inherits configs/base.yaml; keys here override. | |
| extends: "base.yaml" | |
| task: "g" | |
| output_dir: "runs/generator_g" | |
| hub_model_id: null # e.g. "your-username/mathcompose-generator-1.5b" | |
| data: | |
| train_path: "data/generator/train.jsonl" | |
| val_path: "data/generator/val.jsonl" | |
| # jsonl rows are {"prompt": <problem as chat messages or str>, "completion": <CoT + \boxed{answer}>} | |
| train: | |
| max_length: 2048 | |
| per_device_train_batch_size: 2 | |
| gradient_accumulation_steps: 8 | |
| eval: | |
| math500_id: "HuggingFaceH4/MATH-500" | |
| aime_id: "Maxwell-Jia/AIME_2024" | |
| # Composition (V reranks G): | |
| compose_n: 8 # samples per problem | |
| compose_temperature: 0.8 | |