SkillScreen-2B-Markdown / evaluation.json
JerroldH's picture
Release v0.1.0: full-data deployment model, inference entry point and OOF thresholds
b60475f verified
Raw History Blame Contribute Delete
3.5 kB
{
"mode": "markdown8k",
"seed": 42,
"records": 13514,
"malicious": 7500,
"benign": 6014,
"auprc_average_precision": 0.9966339274117415,
"auroc": 0.9945986919410504,
"folds_sha256": "c11c6f68dd23fddb70d7090512d93f5be54bb216596175d44fcaa7e7df80e202",
"oof_sha256": "935e97b51798f6b553da87c4cc0738510a7575c14b0fe4bc6db2726145baa888",
"folds": [
{
"fold": 1,
"predictions_sha256": "a43bd55e2f3256b7e5d959822ce449ecee79dfc2affbb1a3c6a6fab04bbc1ec0",
"adapter_sha256": "79c2cfa9e230b5eb88dc26ae96a5f9b26ea86d6f7f5d187fbb9070b9518d7ec4"
},
{
"fold": 2,
"predictions_sha256": "608d14ae22ea7579928f0722dc4ba9a4a7dfad923a57e8542d7ad7ae132ed52e",
"adapter_sha256": "7683ceca69861929e1263dd46bf3befeb2534592b90686ad60b4abf21b2841dc"
},
{
"fold": 3,
"predictions_sha256": "a759c2e629bf2e21382806eb332e0760de792683b337f2529f019e0b730286ba",
"adapter_sha256": "3110418b5329db127917bad3135db46d97610cf126cc9ced6b08f048b7ba5100"
}
],
"created_at": "2026-11-10T16:05:55.944354+00:00",
"default": {
"raw_threshold": 0.0,
"native_threshold": 0.5,
"tp": 7272,
"fp": 53,
"fn": 228,
"tn": 5961,
"recall": 0.9696,
"fpr": 0.008812770202859994,
"precision": 0.9927645051194539,
"f1": 0.9810455311973019,
"flagged_fraction": 0.5420304869024715
},
"targets": {
"0.95": {
"raw_threshold": 5.374836521935736,
"native_threshold": 0.9953896776695467,
"tp": 7125,
"fp": 9,
"fn": 375,
"tn": 6005,
"recall": 0.95,
"fpr": 0.0014965081476554707,
"precision": 0.9987384356602187,
"f1": 0.973759737597376,
"flagged_fraction": 0.5278969957081545
},
"0.98": {
"raw_threshold": -2.9843786119895848,
"native_threshold": 0.0481366056169061,
"tp": 7350,
"fp": 243,
"fn": 150,
"tn": 5771,
"recall": 0.98,
"fpr": 0.0404057199866977,
"precision": 0.9679968391939945,
"f1": 0.9739614390777181,
"flagged_fraction": 0.5618617729761729
},
"0.99": {
"raw_threshold": -4.441408183360328,
"native_threshold": 0.011642201918123018,
"tp": 7425,
"fp": 945,
"fn": 75,
"tn": 5069,
"recall": 0.99,
"fpr": 0.1571333555038244,
"precision": 0.8870967741935484,
"f1": 0.9357277882797732,
"flagged_fraction": 0.6193577031226876
}
},
"balanced": {
"target": 0.95,
"raw_threshold": 5.374836521935736,
"native_threshold": 0.9953896776695467,
"tp": 7125,
"fp": 9,
"fn": 375,
"tn": 6005,
"recall": 0.95,
"fpr": 0.0014965081476554707,
"precision": 0.9987384356602187,
"f1": 0.973759737597376,
"flagged_fraction": 0.5278969957081545
},
"high_recall": {
"raw_threshold": -4.441408183360328,
"native_threshold": 0.011642201918123018,
"tp": 7425,
"fp": 945,
"fn": 75,
"tn": 5069,
"recall": 0.99,
"fpr": 0.1571333555038244,
"precision": 0.8870967741935484,
"f1": 0.9357277882797732,
"flagged_fraction": 0.6193577031226876
},
"selection": "Balanced maximizes recall minus FPR among predeclared 95/98/99% OOF recall gates; ties prefer higher recall. High-recall targets 99% OOF recall.",
"scope": "Pooled three-fold out-of-fold scores. Every score is from a model that excluded that skill group. Thresholds transfer to a separate full-data refit; these are not held-out test results of that refit."
}