| """Compare predicted decisions and already-computed results; never execute SQL.""" |
| import json |
|
|
| def model_input(row): |
| return {k:row[k] for k in ('database_id','database_path','schema_sql','question_ar')} |
|
|
| def score(row, *, decision, result): |
| if decision not in ('execute','clarify','unanswerable'): |
| raise ValueError('decision must be execute, clarify or unanswerable') |
| if decision == 'execute' and not isinstance(result,list): |
| raise ValueError('execute requires a JSON-style list of rows') |
| if decision in ('clarify','unanswerable') and result is not None: |
| raise ValueError('clarify and unanswerable require result=None') |
| decision_correct=decision==row['decision'] |
| expected=json.loads(row['expected_result_json']) if row['expected_result_json'] is not None else None |
| canonical=lambda v:json.dumps(v,ensure_ascii=False,sort_keys=True,allow_nan=False,separators=(',',':')) |
| return {'decision_correct':decision_correct,'joint_correct':decision_correct and canonical(result)==canonical(expected)} |
|
|
