109 lines
2.3 KiB
JSON
109 lines
2.3 KiB
JSON
{
|
|
"model": "Qwen/Qwen3.6-35B-A3B-FP8",
|
|
"api_base": "http://localhost:8000/v1",
|
|
"timestamp": "2026-05-05T23:48:26.106574",
|
|
"total_evaluated": 60,
|
|
"summary": {
|
|
"total_episodes": 60,
|
|
"detection_accuracy": 0.5666666666666667,
|
|
"miss_rate": 0.6785714285714286,
|
|
"false_alarm_rate": 0.21875,
|
|
"precision": 0.5625,
|
|
"recall": 0.32142857142857145,
|
|
"f1_security": 0.40909090909090917,
|
|
"threat_type_accuracy": 0.3333333333333333,
|
|
"parse_failure_rate": 0.0,
|
|
"breakdown": {
|
|
"TP": {
|
|
"total": 28,
|
|
"correct": 9,
|
|
"accuracy": 0.32142857142857145
|
|
},
|
|
"FP": {
|
|
"total": 28,
|
|
"correct": 21,
|
|
"accuracy": 0.75
|
|
},
|
|
"TN": {
|
|
"total": 4,
|
|
"correct": 4,
|
|
"accuracy": 1.0
|
|
}
|
|
},
|
|
"per_sq": {
|
|
"SQ1": {
|
|
"total": 13,
|
|
"accuracy": 0.5384615384615384,
|
|
"miss_count": 5,
|
|
"false_alarm_count": 1
|
|
},
|
|
"SQ2": {
|
|
"total": 6,
|
|
"accuracy": 0.6666666666666666,
|
|
"miss_count": 0,
|
|
"false_alarm_count": 2
|
|
},
|
|
"SQ3": {
|
|
"total": 17,
|
|
"accuracy": 0.47058823529411764,
|
|
"miss_count": 7,
|
|
"false_alarm_count": 2
|
|
},
|
|
"SQ4": {
|
|
"total": 17,
|
|
"accuracy": 0.6470588235294118,
|
|
"miss_count": 5,
|
|
"false_alarm_count": 1
|
|
},
|
|
"SQ5": {
|
|
"total": 7,
|
|
"accuracy": 0.5714285714285714,
|
|
"miss_count": 2,
|
|
"false_alarm_count": 1
|
|
}
|
|
}
|
|
},
|
|
"errors": {
|
|
"total_errors": 32,
|
|
"error_distribution": {
|
|
"FALSE_ALARM": 7,
|
|
"MISS": 19,
|
|
"WRONG_TYPE": 6
|
|
},
|
|
"error_by_sq": {
|
|
"SQ1": {
|
|
"FALSE_ALARM": 1,
|
|
"MISS": 5
|
|
},
|
|
"SQ2": {
|
|
"FALSE_ALARM": 2,
|
|
"WRONG_TYPE": 2
|
|
},
|
|
"SQ3": {
|
|
"FALSE_ALARM": 2,
|
|
"MISS": 7,
|
|
"WRONG_TYPE": 1
|
|
},
|
|
"SQ4": {
|
|
"FALSE_ALARM": 1,
|
|
"MISS": 5,
|
|
"WRONG_TYPE": 2
|
|
},
|
|
"SQ5": {
|
|
"FALSE_ALARM": 1,
|
|
"MISS": 2,
|
|
"WRONG_TYPE": 1
|
|
}
|
|
},
|
|
"error_by_category": {
|
|
"device_fault": 5,
|
|
"intrusion": 4,
|
|
"water_damage": 2,
|
|
"behavioral_anomaly": 4,
|
|
"child_specific": 4,
|
|
"elderly_specific": 4,
|
|
"fire_gas": 2
|
|
}
|
|
},
|
|
"pipeline": "EGPv4"
|
|
} |