| { | |
| "canonical_answer_nll": 1.252042673200658, | |
| "canonical_answer_perplexity": 3.4974798741948536, | |
| "evaluated_prompts": 2477, | |
| "generation_rouge_l_f1": 0.3024088364317219, | |
| "generation_token_f1": 0.42547394221365986, | |
| "path_samples_per_prompt": 4, | |
| "preparation": { | |
| "discarded_prompts": 111, | |
| "discarded_responses": 111, | |
| "kept_prompts": 2477 | |
| }, | |
| "reading_logprob_std": 1.261352655227725, | |
| "reading_span_available_prompts": 2477, | |
| "reading_span_unavailable_prompts": 0, | |
| "tokenization_path_logprob_std": 114.96431571929517, | |
| "trajectory_balance_rmse": 3209.4460347700096 | |
| } | |