Seter: open models that catch Hebrew and Arabic names

Free, MIT-licensed PII detection for Hebrew, Arabic, English and mixed text. Includes first names on their own, which generic filters miss.

Try it in your browser · Models on Hugging Face · Self-host the guard

The browser demo runs the fast model on your machine (int8 ONNX, Transformers.js), so text never leaves the page. The accurate v1.1 model is the large checkpoint on Hugging Face.

Measured, not claimed

On the REDACT Hebrew, Arabic and English records (LLM-generated — so synthetic-on-synthetic; evaluation only, never training), Seter v1.1 improves on the previous release and on OpenAI’s open privacy filter.

Large-model figures are the mean of 3 training seeds; the released weights are seed 1 (in parentheses). The fast model is still the 25 Oct 2026 v1.0 release. Source: Vats et al., REDACT, arXiv:2606.19881 (2026). Method and caveats are on the model card.

Hebrew first names on their own, on a fixed-seed REDACT subset (250 HE / 250 AR / 150 EN; ≥50% of the name covered by a person span):

Same subset, Arabic / English first-name recall for Seter v1.1: 0.921 / 0.952. Competitor rows are from the 3 Nov 2026 CPU bench (uniform decoder); Seter v1.1 was scored with the shipped shared decoder.

Payment-card recall on the released v1.1 weights, including last-four and masked forms: 98% / 90% / 95% Hebrew / Arabic / English (v1.0: 50% / 38% / 38%). Names caught (any overlap): Hebrew 99.5% vs 40% for openai/privacy-filter.

The honest part: REDACT is LLM-generated. Seter still masks more aggressively than GLiNER or privacy-filter, which means more false positives. Pick it when missing a name costs more than over-masking a word. Not yet compared against Azure or Private AI. The fast model has not been retrained for v1.1.

What you get (all free)

• seter-pii-he-ar — full model (xlm-roberta-large), v1.1, best accuracy.
• seter-pii-he-ar-fast — mmBERT-base, int8 ONNX, runs in the browser (still v1.0).
• seter-guard — Docker REST API plus a LiteLLM guardrail, to mask prompts before they reach any LLM.
• Labels: private_person, private_address, private_phone, private_email, account_number. Same names as openai/privacy-filter.
• Trained only on synthetic data. No real personal data was used.

git clone https://github.com/Shoresh-Labs/seter-guard && cd seter-guard
docker build -t seter-guard . && docker run -p 8080:8080 seter-guard

POST /v1/redact   {"text": "..."}          # or plug into LiteLLM as a guardrail