auto-3b / calibration.json
ProCreations's picture
Release auto-3b (SmolLM3-3B-Base approve/deny classifier, 98.03% on Approve-or-Deny)
58323da verified
Raw History Blame Contribute Delete
711 Bytes
{
"threshold": 0.295,
"validation_metrics": {
"n": 2581,
"accuracy": 0.9833397907787679,
"balanced_accuracy": 0.9833393179247119,
"f1_deny": 0.9833655705996132,
"auroc": 0.9978311360475756,
"nll": 0.06344092496728404,
"false_approve_rate": 0.016253869969040248,
"false_deny_rate": 0.017067494181536073,
"false_approve_count": 21,
"false_deny_count": 22,
"deny_count": 1292,
"approve_count": 1289,
"threshold": 0.295,
"brier": 0.015282586776204823,
"accuracy_ci95": [
0.9776351922332552,
0.9876077597032694
]
},
"criterion": "balanced accuracy; ties prefer fewer false approvals then 0.5 proximity",
"benchmark_used": false
}