109 lines
2.3 KiB
JSON
109 lines
2.3 KiB
JSON
{
|
|
"model": "deepseek-v4-flash",
|
|
"api_base": "https://uni-api.cstcloud.cn/v1",
|
|
"timestamp": "2026-05-04T21:33:57.389834",
|
|
"total_evaluated": 60,
|
|
"summary": {
|
|
"total_episodes": 60,
|
|
"detection_accuracy": 0.5666666666666667,
|
|
"miss_rate": 0.5357142857142857,
|
|
"false_alarm_rate": 0.34375,
|
|
"precision": 0.5416666666666666,
|
|
"recall": 0.4642857142857143,
|
|
"f1_security": 0.5,
|
|
"threat_type_accuracy": 0.3076923076923077,
|
|
"parse_failure_rate": 0.016666666666666666,
|
|
"breakdown": {
|
|
"TP": {
|
|
"total": 28,
|
|
"correct": 13,
|
|
"accuracy": 0.4642857142857143
|
|
},
|
|
"FP": {
|
|
"total": 28,
|
|
"correct": 19,
|
|
"accuracy": 0.6785714285714286
|
|
},
|
|
"TN": {
|
|
"total": 4,
|
|
"correct": 2,
|
|
"accuracy": 0.5
|
|
}
|
|
},
|
|
"per_sq": {
|
|
"SQ1": {
|
|
"total": 13,
|
|
"accuracy": 0.46153846153846156,
|
|
"miss_count": 5,
|
|
"false_alarm_count": 2
|
|
},
|
|
"SQ2": {
|
|
"total": 6,
|
|
"accuracy": 0.3333333333333333,
|
|
"miss_count": 2,
|
|
"false_alarm_count": 2
|
|
},
|
|
"SQ3": {
|
|
"total": 17,
|
|
"accuracy": 0.5882352941176471,
|
|
"miss_count": 4,
|
|
"false_alarm_count": 3
|
|
},
|
|
"SQ4": {
|
|
"total": 17,
|
|
"accuracy": 0.6470588235294118,
|
|
"miss_count": 2,
|
|
"false_alarm_count": 4
|
|
},
|
|
"SQ5": {
|
|
"total": 7,
|
|
"accuracy": 0.7142857142857143,
|
|
"miss_count": 2,
|
|
"false_alarm_count": 0
|
|
}
|
|
}
|
|
},
|
|
"errors": {
|
|
"total_errors": 35,
|
|
"error_distribution": {
|
|
"FALSE_ALARM": 11,
|
|
"MISS": 14,
|
|
"PARSE_FAIL": 1,
|
|
"WRONG_TYPE": 9
|
|
},
|
|
"error_by_sq": {
|
|
"SQ1": {
|
|
"FALSE_ALARM": 2,
|
|
"MISS": 5
|
|
},
|
|
"SQ2": {
|
|
"FALSE_ALARM": 2,
|
|
"MISS": 2
|
|
},
|
|
"SQ3": {
|
|
"FALSE_ALARM": 3,
|
|
"PARSE_FAIL": 1,
|
|
"WRONG_TYPE": 4,
|
|
"MISS": 3
|
|
},
|
|
"SQ4": {
|
|
"FALSE_ALARM": 4,
|
|
"MISS": 2,
|
|
"WRONG_TYPE": 5
|
|
},
|
|
"SQ5": {
|
|
"MISS": 2
|
|
}
|
|
},
|
|
"error_by_category": {
|
|
"device_fault": 5,
|
|
"intrusion": 3,
|
|
"water_damage": 2,
|
|
"behavioral_anomaly": 4,
|
|
"child_specific": 4,
|
|
"elderly_specific": 4,
|
|
"fire_gas": 1
|
|
}
|
|
},
|
|
"pipeline": "EGPv2"
|
|
} |