Ettin encoder 32M fine-tuned on Nemotron-PII (ONNX)

English PII named-entity model used by redact-web to find names, addresses and other personal information in text, in the browser.

• Upstream: https://huggingface.co/kalyan-ks/ettin-32m-nemotron-pii, at commit 4d2ce98244262fe1461a834d4b2ce653df7009a4 (MIT, Kalyan KS).
• Base model: jhu-clsp/ettin-encoder-32m (MIT, Johns Hopkins CLSP, arXiv:2507.11412).
• Fine-tuning data: NVIDIA Nemotron-PII, "Nemotron-PII: Synthesized Data for Privacy-Preserving AI", https://huggingface.co/datasets/nvidia/Nemotron-PII, licensed CC BY 4.0. Attribution: NVIDIA Corporation.
• License: MIT, the same as upstream.

Our modifications

Built by tools/models/ner/export.py:

• Exported from PyTorch to ONNX (input_ids, attention_mask int64 [B, S] to logits float32 [B, S, 107]), with parity checked against PyTorch at 31 tokens, at 512 tokens and on a padded batch.
• An fp16 copy that keeps float32 inputs and outputs.
• tokenizer.json and config.json are copied unchanged from upstream; redact-web reads the labels from config.json (id2label).

Files

redact-web pins this repository at a commit and verifies every file's SHA-256 before use.