| { | |
| "overall_accuracy": 0.412, | |
| "highest_tier_above_90": 4, | |
| "deterministic": true, | |
| "tiers": [ | |
| { | |
| "tier_id": 0, | |
| "total": 100, | |
| "correct": 20, | |
| "accuracy": 0.2, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 1, | |
| "total": 100, | |
| "correct": 100, | |
| "accuracy": 1.0, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 2, | |
| "total": 100, | |
| "correct": 100, | |
| "accuracy": 1.0, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 3, | |
| "total": 100, | |
| "correct": 100, | |
| "accuracy": 1.0, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 4, | |
| "total": 100, | |
| "correct": 100, | |
| "accuracy": 1.0, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 5, | |
| "total": 100, | |
| "correct": 2, | |
| "accuracy": 0.02, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 6, | |
| "total": 100, | |
| "correct": 2, | |
| "accuracy": 0.02, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 7, | |
| "total": 100, | |
| "correct": 2, | |
| "accuracy": 0.02, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 8, | |
| "total": 100, | |
| "correct": 2, | |
| "accuracy": 0.02, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 9, | |
| "total": 100, | |
| "correct": 2, | |
| "accuracy": 0.02, | |
| "completed": true | |
| }, | |
| { | |
| "tier_id": 10, | |
| "total": 100, | |
| "correct": 2, | |
| "accuracy": 0.02, | |
| "completed": true | |
| } | |
| ], | |
| "repo_id": "", | |
| "revision": "", | |
| "eval_period": "" | |
| } | |