| { |
| "train": { |
| "readout": { |
| "categorical": "175/246", |
| "categorical_accuracy": 0.7113821138211383, |
| "completion_found": "45/48", |
| "false_completions": "1/96", |
| "unknown": "0/48", |
| "score_mae": 0.39696706232747864 |
| }, |
| "verify_zero_shot": { |
| "categorical": "177/246", |
| "categorical_accuracy": 0.7195121951219512, |
| "completion_found": "47/48", |
| "false_completions": "0/96", |
| "unknown": "0/48", |
| "score_mae": 0.37552783307934967 |
| }, |
| "seconds_per_request": { |
| "mean": 1.3231873131460614, |
| "p95": 1.7301512956619263 |
| } |
| }, |
| "validation": { |
| "readout": { |
| "categorical": "92/126", |
| "categorical_accuracy": 0.7301587301587301, |
| "completion_found": "23/24", |
| "false_completions": "0/48", |
| "unknown": "2/24", |
| "score_mae": 0.3990468073946734 |
| }, |
| "verify_zero_shot": { |
| "categorical": "89/126", |
| "categorical_accuracy": 0.7063492063492064, |
| "completion_found": "21/24", |
| "false_completions": "0/48", |
| "unknown": "1/24", |
| "score_mae": 0.28145276563851634 |
| }, |
| "seconds_per_request": { |
| "mean": 1.2563178936640422, |
| "p95": 1.640113389492035 |
| } |
| }, |
| "test": { |
| "readout": { |
| "categorical": "86/126", |
| "categorical_accuracy": 0.6825396825396826, |
| "completion_found": "24/24", |
| "false_completions": "0/48", |
| "unknown": "0/24", |
| "score_mae": 0.5369042108585644 |
| }, |
| "verify_zero_shot": { |
| "categorical": "85/126", |
| "categorical_accuracy": 0.6746031746031746, |
| "completion_found": "22/24", |
| "false_completions": "0/48", |
| "unknown": "0/24", |
| "score_mae": 0.42307032748062895 |
| }, |
| "seconds_per_request": { |
| "mean": 1.279320968521966, |
| "p95": 1.6923456430435182 |
| } |
| }, |
| "total_minutes": 3.8 |
| } |
