| { |
| "canonical_answer_nll": 1.352831203965315, |
| "canonical_answer_perplexity": 3.868362163769441, |
| "evaluated_prompts": 2477, |
| "generation_rouge_l_f1": 0.29973091403228963, |
| "generation_token_f1": 0.42380786949584803, |
| "path_samples_per_prompt": 4, |
| "preparation": { |
| "discarded_prompts": 111, |
| "discarded_responses": 111, |
| "kept_prompts": 2477 |
| }, |
| "reading_logprob_std": 1.9123915398505373, |
| "reading_span_available_prompts": 2477, |
| "reading_span_unavailable_prompts": 0, |
| "tokenization_path_logprob_std": 162.5901912800582, |
| "trajectory_balance_rmse": 4440.580532213749 |
| } |
|
|
