ege-arhan/TrustLaya-S-Advanced

TrustLaya-S Advanced: experimental v2 PII candidate

42,138,641 parameter Turkish-first multitask encoder. This version updates only the PII head of TrustLaya-S v1. It adds no new teacher distillation. A separate deterministic policy engine, agent/session risk rules, and evidence extraction live in the source branch. The model alone does not implement the full policy system.

Research candidate only. Not production-ready; do not use as sole gate for irreversible agent actions, personal-data disclosure, or legal/ethical decisions. Risk scores are task-model outputs, not validated real-world event probabilities. confidence is not calibrated correctness.

Independent evaluation

A separate CC-BY 4.0 synthetic Turkish PII test sample (n=2,000; 1,000 task positives and 1,000 task-specific negatives) gave v2 hybrid PII F1 0.784, recall 0.848, and false-positive rate 0.316 at a development-selected threshold of 0.8. This high false-positive rate blocks deployment. On the original synthetic mixed-only 1,975-row test, macro F1 0.663, security F1 0.777, injection F1 0.555, data-governance F1 0.000. Encoded prompt injection F1 0.000 on a small controlled suite. The 33-case language smoke suite is too small to establish multilingual performance.

Original English-only synthetic validation mean raw/calibrated ECE 0.105/0.089, Brier 0.103/0.085, NLL 0.424/0.271. Eight temperatures were evaluated on the same data used to fit them. Turkish PII test ECE 0.181. More detail: research report, model card, data card.

Firewall runtime demo (unchanged weights)

The same trustlaya_s.onnx now runs inside a Docker firewall demo on the feat/v2-firewall-e2e branch: agent → gateway (this model + policy + single-use authorization) → trusted adapter → SQLite record.write target, on internal networks, with no torch in the runtime image. Analysis responses now report reading coverage: the model reads only the first 94 content tokens (head_94_v1), and an unread remainder turns ALLOW/REDACT into REVIEW for protected actions. Targets are idempotent per operation_id; a lost target reply is reported as unknown, not "not executed". See the demo documentation.

All 13 fixed scenarios passed on 2026-09-26 (Docker on Apple Silicon): no unauthorized scenario created a target record, a retried operation never created a second record, and no key or secret appeared in logs. End-to-end p95 53 ms, model p95 45 ms over 100 sequential writes; gateway memory about 270 MiB. These are fixed demonstration inputs, not a quality benchmark. Observed model limits in the same runs: many short benign notes received REVIEW, and a 113-token benign meeting note was BLOCKed as prompt injection. No UNO Q measurement.

Public human-data baseline (2026-09-26, unchanged weights)

The frozen V2 prompt-injection head was scored at its native 94-token reading and a threshold of 0.5 fixed in advance, on cluster-deduplicated public data (V5 evidence gate). Tensor Trust TEST attacks (n=3,596, attack-only): recall 0.443. JailbreakLLMs TEST (172 jailbreak / 1,938 regular prompts): recall 0.901, false-positive rate 0.875, F1 0.153. Benign security prose: false-positive rate 0.898 (documentation/forum) and 0.978 (arXiv abstracts). deepset (OOD): F1 0.710, FPR 0.364. Gandalf (OOD, attack-only): recall 0.722. These results show V2 is not usable as a general jailbreak/injection detector on real-world text. A first V5 training ablation (8 runs on these public sources) was NO_GO: in-source results improved, but false positives on an independent benign set (deepset) rose to 0.70–0.85, so this V2 checkpoint stays the published model (ablation report).

Files and usage

model.safetensors is a custom nine-risk-head PyTorch model, not a generic AutoModel. trustlaya_s.onnx is FP32; trustlaya_s_int8.onnx is experimental and changed 5.1% of final policy actions on 256 synthetic rows. calibration.json, decision_thresholds.json, policy.yaml, and tokenizer files support the full pipeline. Batch-1 MacBook ONNX CPU p50 4.629 ms in the latest run; no Arduino UNO Q hardware benchmark was performed.

Use the source code and its Analyzer with this model directory and an explicit ONNX path. Provenance: MIT YTU Turkish BERT backbone; weak v1 teacher was Apache-2.0 Laya Multilingual, never ground truth. V2 PII training source is MIT Turkish Privacy Filter Dataset. No source examples are redistributed here.