curtain-privacy

On-device PII redaction: a token classifier plus deterministic recognizers across 24 Latin-script languages, with a reversible-placeholder layer for chat. It runs in the browser, in Node, and on iOS and Android from these same files. Source, training pipeline, whitepaper, and the full model card: https://github.com/hackshare/curtain-privacy

This repo ships in two sizes, selected by revision:

• curtain-small — v2.0.0, 59.9 MB, 12-layer Multilingual-MiniLM with a SentencePiece tokenizer. The default: best multilingual and repeated/leading-name recall. onnx/model_q4.onnx SHA-256 c9bfe8a0a9e3cfbb1a8995009259a5195109c8881f16905ef5be58351ad8786f.
• curtain-tiny — v1.0.0, 14.2 MB, 6-layer MiniLM with a WordPiece tokenizer. The smallest-footprint tier, frozen. onnx/model_q4.onnx SHA-256 24ba1f03a8c3db8a4f760d4d266faeb679ba23f8c23f7be9ba6964cbfff6f6c1.

The models are built locally from the committed train/ pipeline; this repo is a distribution copy. Verify integrity by pinning a revision and checking the hash above.

Integrating

Web and Node (transformers.js)

import { createGuard } from "curtain-privacy";

const guard = await createGuard({ model: "hackshare/curtain-privacy", revision: "v2.0.0" });
const { text } = await guard.protect("My SSN is 472-81-0094");

revision: "v2.0.0" pins the audited curtain-small weights (use v1.0.0 for curtain-tiny). transformers.js fetches from this repo with permissive CORS, so it works from any origin. v2.0.0 is a moving pointer to the latest curtain-small build; to pin an immutable artifact, use the timestamped tag v2.0.0-<timestamp> published alongside each release.

iOS and Android (ONNX Runtime Mobile)

Fetch or bundle two files from a pinned revision (curtain-small shown; use v1.0.0 for curtain-tiny):

• https://huggingface.co/hackshare/curtain-privacy/resolve/v2.0.0/onnx/model_q4.onnx
• https://huggingface.co/hackshare/curtain-privacy/resolve/v2.0.0/tokenizer.json

Verify model_q4.onnx against the SHA-256 above. Load the model with ONNX Runtime Mobile, tokenize with the Hugging Face tokenizers library (Swift and Kotlin bindings load tokenizer.json directly) or ONNX Runtime Extensions, then apply the BIO decode and the default keep-set (CITY, STATE, ZIP_CODE) the library documents. These are integration guides, not a shipped native SDK.