Open PII models for Arabic: English-first filters are weak on names

Seter, an MIT-licensed PII model for Arabic, Hebrew and English released by Shoresh Labs, reaches a lenient character F1 of 0.917 on Arabic, 0.904 on Hebrew and 0.878 on English, and catches 88% of standalone Hebrew first names where openai/privacy-filter caught 14% and GLiNER 33%. The model runs fully in the browser via ONNX, and its author says English-first PII filters are weak on Arabic names, including standalone first names and kunyas. The release matters because off-the-shelf redaction tools did not hold up on Semitic names in a synthetic stress test of standalone Hebrew first names.

English-first PII filters are weak on Arabic names, including standalone first names and kunyas. I ran into this while redacting text before sending it to an LLM. Off-the-shelf tools did not hold up on Semitic names more generally. On a synthetic set of standalone Hebrew first names a close stress test for the same class of errors , openai/privacy-filter caught 14% and GLiNER 33%. I trained Seter, an open MIT PII model for Arabic, Hebrew and English. Lenient char-F1 is 0.917 on Arabic, 0.904 on Hebrew and 0.878 on English. On that same first-name set it catches 88%. It runs fully in the browser via ONNX. Demo: Seter — Hebrew/Arabic PII redaction - a Hugging Face Space by yinonov https://huggingface.co/spaces/yinonov/seter-demo Model: shoresh-labs/seter-pii-he-ar · Hugging Face https://huggingface.co/shoresh-labs/seter-pii-he-ar Failure cases from real-world Arabic text are the most useful thing you could send back.