| experiment: | |
| project: "your_project" # Output directory for evaluation results | |
| rollout: | |
| tensor_parallel_size: 1 # set to 1 by default, if oom, try reduce max_active first, if still oom, set tensor_parallel_size to 8 | |
| max_active: 256 | |
| num_task_per_step: 1 | |
| num_response_per_task: 1 | |
| temperature: 0.1 | |
| max_token: 2000 | |
| block_size: 4 | |
| denoising_steps_per_block: 4 | |
| top_p: 1.0 | |
| top_k: 0 | |
| remasking_strategy: "low_confidence_static" #"low_confidence_static""low_confidence_dynamic" | |
| dynamic_threshold: 0.9 # no use for "low_confidence_static" | |
| start_with_think: False | |
| random_init_ratio: 0.0 | |
| # Batch processing: split dataset into batches to avoid KV cache exhaustion | |
| # If None or 0, process all prompts at once (original behavior) | |
| # If set to a positive number, process prompts in batches of this size | |
| batch_size: 50 # e.g., set to 50 to process 50 prompts at a time in case we run out of KV cache space | |
| evaluation: | |
| # Model checkpoint to evaluate (overrides 'model' above if specified) | |
| checkpoint_path: "/path/to/your/checkpoint" | |
| # Dataset configuration | |
| # eval_dataset: "MATH_train" # Dataset name (must exist in data/{eval_dataset}.json) | |
| eval_dataset: "your_dataset" | |
| data_type: "math" # Task type: "math", "code", or "option" | |
| # Distributed evaluation (for multi-node setups) | |
| num_node: 1 # Total number of nodes | |
| node_index: 0 # Current node index (0-indexed) | |
| # Output options | |
| output_unmasking_history: true # Save step-by-step unmasking history | |