Files
llmiotsafe/results/qwen35_egpv2_diag60_after_fix/summary.json
2026-05-12 17:01:39 +08:00

101 lines
2.1 KiB
JSON

{
"model": "Qwen/Qwen3.5-9B",
"api_base": "http://localhost:8000/v1",
"timestamp": "2026-05-04T16:45:14.601677",
"total_evaluated": 60,
"summary": {
"total_episodes": 60,
"detection_accuracy": 0.5333333333333333,
"miss_rate": 0.9642857142857143,
"false_alarm_rate": 0.03125,
"precision": 0.5,
"recall": 0.03571428571428571,
"f1_security": 0.06666666666666667,
"threat_type_accuracy": 1.0,
"parse_failure_rate": 0.0,
"breakdown": {
"TP": {
"total": 28,
"correct": 1,
"accuracy": 0.03571428571428571
},
"FP": {
"total": 28,
"correct": 27,
"accuracy": 0.9642857142857143
},
"TN": {
"total": 4,
"correct": 4,
"accuracy": 1.0
}
},
"per_sq": {
"SQ1": {
"total": 13,
"accuracy": 0.6153846153846154,
"miss_count": 5,
"false_alarm_count": 0
},
"SQ2": {
"total": 6,
"accuracy": 0.5,
"miss_count": 3,
"false_alarm_count": 0
},
"SQ3": {
"total": 17,
"accuracy": 0.5294117647058824,
"miss_count": 8,
"false_alarm_count": 0
},
"SQ4": {
"total": 17,
"accuracy": 0.47058823529411764,
"miss_count": 8,
"false_alarm_count": 1
},
"SQ5": {
"total": 7,
"accuracy": 0.5714285714285714,
"miss_count": 3,
"false_alarm_count": 0
}
}
},
"errors": {
"total_errors": 28,
"error_distribution": {
"MISS": 27,
"FALSE_ALARM": 1
},
"error_by_sq": {
"SQ1": {
"MISS": 5
},
"SQ2": {
"MISS": 3
},
"SQ3": {
"MISS": 8
},
"SQ4": {
"FALSE_ALARM": 1,
"MISS": 8
},
"SQ5": {
"MISS": 3
}
},
"error_by_category": {
"device_fault": 5,
"water_damage": 2,
"intrusion": 4,
"fire_gas": 4,
"behavioral_anomaly": 4,
"child_specific": 4,
"elderly_specific": 4
}
},
"pipeline": "EGPv2"
}