| { |
| "schema_analysis": { |
| "metrics": { |
| "f1_score": 0.8089157550696012, |
| "precision": 0.8324027324027324, |
| "recall": 0.8083877833877834 |
| }, |
| "num_underpredictions": 210, |
| "num_overpredictions": 112, |
| "num_exact_schema_matches": 769, |
| "avg_dist_gold_serialized": 0.81, |
| "avg_dist_gold_predicted": 0.93, |
| "num_hallucinated_schemas": 59 |
| }, |
| "sql_analysis": { |
| "exec": 0.5686070686070686, |
| "em": 0.5343035343035343, |
| "malformed_pred_sql": 119, |
| "correct_over_t5_large": 62, |
| "items_with_counting_errors": 74, |
| "edit_up": 0.1083, |
| "edit_down": 0.7536, |
| "progress": 0.562877905848566 |
| } |
| } |
