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
.
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.
TrustLaya-S-Advanced huggingface.co is an AI model on huggingface.co that provides TrustLaya-S-Advanced's model effect (), which can be used instantly with this ege-arhan TrustLaya-S-Advanced model. huggingface.co supports a free trial of the TrustLaya-S-Advanced model, and also provides paid use of the TrustLaya-S-Advanced. Support call TrustLaya-S-Advanced model through api, including Node.js, Python, http.
TrustLaya-S-Advanced huggingface.co is an online trial and call api platform, which integrates TrustLaya-S-Advanced's modeling effects, including api services, and provides a free online trial of TrustLaya-S-Advanced, you can try TrustLaya-S-Advanced online for free by clicking the link below.
ege-arhan TrustLaya-S-Advanced online free url in huggingface.co:
TrustLaya-S-Advanced is an open source model from GitHub that offers a free installation service, and any user can find TrustLaya-S-Advanced on GitHub to install. At the same time, huggingface.co provides the effect of TrustLaya-S-Advanced install, users can directly use TrustLaya-S-Advanced installed effect in huggingface.co for debugging and trial. It also supports api for free installation.
TrustLaya-S-Advanced install url in huggingface.co:
