data-quality-triage-assistant / scripts /trajectory_eval_results.json
dhrumilparikh's picture
Upload folder using huggingface_hub
f30bceb verified
Raw
History Blame Contribute Delete
13.8 kB
{
"results": {
"easy_missing_and_dupes": {
"good_policy": {
"cumulative_reward": 0.36,
"final_score": 0.24444444444444444,
"evaluation": {
"schema_version": "1.0.0",
"decision": "approved",
"gates": [
{
"name": "min_final_score",
"passed": true,
"actual": 0.24444444444444444,
"expected": 0.2,
"comparator": ">=",
"reason": "Ensures task performance quality."
},
{
"name": "max_invalid_actions",
"passed": true,
"actual": 0.0,
"expected": 1.0,
"comparator": "<=",
"reason": "Limits invalid operational behavior."
},
{
"name": "max_risk_score",
"passed": true,
"actual": 45.0,
"expected": 70.0,
"comparator": "<=",
"reason": "Controls governance risk ceiling."
},
{
"name": "min_issue_reduction_ratio",
"passed": true,
"actual": 0.4444444444444444,
"expected": 0.35,
"comparator": ">=",
"reason": "Requires meaningful data quality improvement."
}
],
"failed_gates": [],
"thresholds": {
"min_final_score": 0.2,
"max_invalid_actions": 1.0,
"max_risk_score": 70.0,
"min_issue_reduction_ratio": 0.35
},
"metrics": {
"final_score": 0.24444444444444444,
"issue_reduction_ratio": 0.4444444444444444,
"max_risk_score": 45.0,
"invalid_action_count": 0.0,
"composite_score": 0.315
},
"leaderboard_record": {
"task_id": "easy_missing_and_dupes",
"difficulty": "easy",
"final_score": 0.24444444444444444,
"composite_score": 0.315,
"decision": "approved",
"max_risk_score": 45.0,
"invalid_action_count": 0.0
}
}
},
"bad_policy": {
"cumulative_reward": -0.42,
"final_score": 0.06666666666666668,
"evaluation": {
"schema_version": "1.0.0",
"decision": "rejected",
"gates": [
{
"name": "min_final_score",
"passed": false,
"actual": 0.06666666666666668,
"expected": 0.2,
"comparator": ">=",
"reason": "Ensures task performance quality."
},
{
"name": "max_invalid_actions",
"passed": false,
"actual": 5.0,
"expected": 1.0,
"comparator": "<=",
"reason": "Limits invalid operational behavior."
},
{
"name": "max_risk_score",
"passed": true,
"actual": 55.0,
"expected": 70.0,
"comparator": "<=",
"reason": "Controls governance risk ceiling."
},
{
"name": "min_issue_reduction_ratio",
"passed": false,
"actual": 0.0,
"expected": 0.35,
"comparator": ">=",
"reason": "Requires meaningful data quality improvement."
}
],
"failed_gates": [
"min_final_score",
"max_invalid_actions",
"min_issue_reduction_ratio"
],
"thresholds": {
"min_final_score": 0.2,
"max_invalid_actions": 1.0,
"max_risk_score": 70.0,
"min_issue_reduction_ratio": 0.35
},
"metrics": {
"final_score": 0.06666666666666668,
"issue_reduction_ratio": 0.0,
"max_risk_score": 55.0,
"invalid_action_count": 5.0,
"composite_score": 0.09166666666666667
},
"leaderboard_record": {
"task_id": "easy_missing_and_dupes",
"difficulty": "easy",
"final_score": 0.06666666666666668,
"composite_score": 0.09166666666666667,
"decision": "rejected",
"max_risk_score": 55.0,
"invalid_action_count": 5.0
}
}
}
},
"medium_type_and_category": {
"good_policy": {
"cumulative_reward": 1.0399999999999998,
"final_score": 0.325974025974026,
"evaluation": {
"schema_version": "1.0.0",
"decision": "approved",
"gates": [
{
"name": "min_final_score",
"passed": true,
"actual": 0.325974025974026,
"expected": 0.26,
"comparator": ">=",
"reason": "Ensures task performance quality."
},
{
"name": "max_invalid_actions",
"passed": true,
"actual": 0.0,
"expected": 1.0,
"comparator": "<=",
"reason": "Limits invalid operational behavior."
},
{
"name": "max_risk_score",
"passed": true,
"actual": 45.0,
"expected": 65.0,
"comparator": "<=",
"reason": "Controls governance risk ceiling."
},
{
"name": "min_issue_reduction_ratio",
"passed": true,
"actual": 0.5277777777777778,
"expected": 0.3,
"comparator": ">=",
"reason": "Requires meaningful data quality improvement."
}
],
"failed_gates": [],
"thresholds": {
"min_final_score": 0.26,
"max_invalid_actions": 1.0,
"max_risk_score": 65.0,
"min_issue_reduction_ratio": 0.3
},
"metrics": {
"final_score": 0.325974025974026,
"issue_reduction_ratio": 0.5277777777777778,
"max_risk_score": 45.0,
"invalid_action_count": 0.0,
"composite_score": 0.38873737373737377
},
"leaderboard_record": {
"task_id": "medium_type_and_category",
"difficulty": "medium",
"final_score": 0.325974025974026,
"composite_score": 0.38873737373737377,
"decision": "approved",
"max_risk_score": 45.0,
"invalid_action_count": 0.0
}
}
},
"bad_policy": {
"cumulative_reward": -0.41000000000000003,
"final_score": 0.09090909090909093,
"evaluation": {
"schema_version": "1.0.0",
"decision": "rejected",
"gates": [
{
"name": "min_final_score",
"passed": false,
"actual": 0.09090909090909093,
"expected": 0.26,
"comparator": ">=",
"reason": "Ensures task performance quality."
},
{
"name": "max_invalid_actions",
"passed": false,
"actual": 5.0,
"expected": 1.0,
"comparator": "<=",
"reason": "Limits invalid operational behavior."
},
{
"name": "max_risk_score",
"passed": true,
"actual": 55.0,
"expected": 65.0,
"comparator": "<=",
"reason": "Controls governance risk ceiling."
},
{
"name": "min_issue_reduction_ratio",
"passed": false,
"actual": 0.0,
"expected": 0.3,
"comparator": ">=",
"reason": "Requires meaningful data quality improvement."
}
],
"failed_gates": [
"min_final_score",
"max_invalid_actions",
"min_issue_reduction_ratio"
],
"thresholds": {
"min_final_score": 0.26,
"max_invalid_actions": 1.0,
"max_risk_score": 65.0,
"min_issue_reduction_ratio": 0.3
},
"metrics": {
"final_score": 0.09090909090909093,
"issue_reduction_ratio": 0.0,
"max_risk_score": 55.0,
"invalid_action_count": 5.0,
"composite_score": 0.10863636363636364
},
"leaderboard_record": {
"task_id": "medium_type_and_category",
"difficulty": "medium",
"final_score": 0.09090909090909093,
"composite_score": 0.10863636363636364,
"decision": "rejected",
"max_risk_score": 55.0,
"invalid_action_count": 5.0
}
}
}
},
"hard_conflicts_and_budget": {
"good_policy": {
"cumulative_reward": 1.03,
"final_score": 0.20379310344827584,
"evaluation": {
"schema_version": "1.0.0",
"decision": "approved",
"gates": [
{
"name": "min_final_score",
"passed": true,
"actual": 0.20379310344827584,
"expected": 0.18,
"comparator": ">=",
"reason": "Ensures task performance quality."
},
{
"name": "max_invalid_actions",
"passed": true,
"actual": 0.0,
"expected": 2.0,
"comparator": "<=",
"reason": "Limits invalid operational behavior."
},
{
"name": "max_risk_score",
"passed": true,
"actual": 45.0,
"expected": 70.0,
"comparator": "<=",
"reason": "Controls governance risk ceiling."
},
{
"name": "min_issue_reduction_ratio",
"passed": true,
"actual": 0.3114754098360656,
"expected": 0.25,
"comparator": ">=",
"reason": "Requires meaningful data quality improvement."
}
],
"failed_gates": [],
"thresholds": {
"min_final_score": 0.18,
"max_invalid_actions": 2.0,
"max_risk_score": 70.0,
"min_issue_reduction_ratio": 0.25
},
"metrics": {
"final_score": 0.20379310344827584,
"issue_reduction_ratio": 0.3114754098360656,
"max_risk_score": 45.0,
"invalid_action_count": 0.0,
"composite_score": 0.2599502543810062
},
"leaderboard_record": {
"task_id": "hard_conflicts_and_budget",
"difficulty": "hard",
"final_score": 0.20379310344827584,
"composite_score": 0.2599502543810062,
"decision": "approved",
"max_risk_score": 45.0,
"invalid_action_count": 0.0
}
}
},
"bad_policy": {
"cumulative_reward": -0.41000000000000003,
"final_score": 0.08000000000000002,
"evaluation": {
"schema_version": "1.0.0",
"decision": "rejected",
"gates": [
{
"name": "min_final_score",
"passed": false,
"actual": 0.08000000000000002,
"expected": 0.18,
"comparator": ">=",
"reason": "Ensures task performance quality."
},
{
"name": "max_invalid_actions",
"passed": false,
"actual": 5.0,
"expected": 2.0,
"comparator": "<=",
"reason": "Limits invalid operational behavior."
},
{
"name": "max_risk_score",
"passed": true,
"actual": 55.0,
"expected": 70.0,
"comparator": "<=",
"reason": "Controls governance risk ceiling."
},
{
"name": "min_issue_reduction_ratio",
"passed": false,
"actual": 0.0,
"expected": 0.25,
"comparator": ">=",
"reason": "Requires meaningful data quality improvement."
}
],
"failed_gates": [
"min_final_score",
"max_invalid_actions",
"min_issue_reduction_ratio"
],
"thresholds": {
"min_final_score": 0.18,
"max_invalid_actions": 2.0,
"max_risk_score": 70.0,
"min_issue_reduction_ratio": 0.25
},
"metrics": {
"final_score": 0.08000000000000002,
"issue_reduction_ratio": 0.0,
"max_risk_score": 55.0,
"invalid_action_count": 5.0,
"composite_score": 0.101
},
"leaderboard_record": {
"task_id": "hard_conflicts_and_budget",
"difficulty": "hard",
"final_score": 0.08000000000000002,
"composite_score": 0.101,
"decision": "rejected",
"max_risk_score": 55.0,
"invalid_action_count": 5.0
}
}
}
}
}
}