This is an experimental fine-tune of Laya for ranking SOC alerts as true-positive vs benign. It was trained on lab/test Wazuh telemetry, not on customer data or customer dispositions.
Not a production SOC model. The model-only / auto-close gate is NO-GO. Labels are a content rubric, not independent analyst dispositions. Use this checkpoint for research, reproduction, and evaluation — not for autonomous alert closure.
typed-decisions (convaiinnovations/laya).noul probability of true_positive; threshold 0.99 for evaluation.model.safetensors).| Metric | Value |
|---|---|
| Held-out rows | 308 (97 positive / 211 negative) |
| Rule-disjoint AUROC | 1.000 |
| Recall at threshold 0.99 | 1.000 |
| FPR at threshold 0.99 | 0.000 |
| Precision at threshold 0.99 | 1.000 |
| Raw ECE | 0.245 (checkpoint temperature noul=0.5; no separate calibration file) |
Do not read these as production accuracy. The labels are rule-derived content rubric; a random split reaches 100% accuracy by rule leakage, and a time split has zero positives in the evaluation period (positive lab alerts only 18–20 Oct). Independent analyst dispositions and a positive time-split are still required.
NOTICE.The bounded policy/configuration and aggregate metrics are published separately: https://huggingface.co/infosecn1nja/laya-soc-bounded-triage
noul=0.5. Raw ECE 0.245.convaiinnovations/laya