60 lines
1.1 KiB
JSON
60 lines
1.1 KiB
JSON
{
|
|
"model": "deepseek-v4-flash",
|
|
"api_base": "https://api.deepseek.com/v1",
|
|
"timestamp": "2026-04-30T17:29:55.646439",
|
|
"total_evaluated": 10,
|
|
"avg_latency_seconds": 17.51,
|
|
"api_errors": 0,
|
|
"summary": {
|
|
"total_episodes": 10,
|
|
"detection_accuracy": 0.8,
|
|
"miss_rate": 0,
|
|
"false_alarm_rate": 0.2,
|
|
"precision": 0.0,
|
|
"recall": 0,
|
|
"f1_security": 0,
|
|
"threat_type_accuracy": 0,
|
|
"parse_failure_rate": 0.0,
|
|
"breakdown": {
|
|
"TP": {
|
|
"total": 0,
|
|
"correct": 0,
|
|
"accuracy": 0
|
|
},
|
|
"FP": {
|
|
"total": 10,
|
|
"correct": 8,
|
|
"accuracy": 0.8
|
|
},
|
|
"TN": {
|
|
"total": 0,
|
|
"correct": 0,
|
|
"accuracy": 0
|
|
}
|
|
},
|
|
"per_sq": {
|
|
"SQ1": {
|
|
"total": 10,
|
|
"accuracy": 0.8,
|
|
"miss_count": 0,
|
|
"false_alarm_count": 2
|
|
}
|
|
}
|
|
},
|
|
"errors": {
|
|
"total_errors": 2,
|
|
"error_distribution": {
|
|
"FALSE_ALARM": 2
|
|
},
|
|
"error_by_sq": {
|
|
"SQ1": {
|
|
"FALSE_ALARM": 2
|
|
},
|
|
"SQ2": {},
|
|
"SQ3": {},
|
|
"SQ4": {},
|
|
"SQ5": {}
|
|
},
|
|
"error_by_category": {}
|
|
}
|
|
} |