PII-Tracer, ONNX 8-bit

An ONNX version of perplexity-ai/PII-Tracer (MIT, Perplexity AI), made for Efitra, a private AI app that runs on your own computer. Efitra uses it in Private mode to find personal details and hide them before a question goes to a cloud model.

This is not a new model: same weights, exported and shrunk.

What was done

  1. Exported from PyTorch to ONNX (opset 18). The fp32 ONNX gives the same result as PyTorch fp32 on all 103 test texts.
  2. Weights quantized to 8 bits in blocks of 32 (MatMulNBits), maths in fp32.
  3. The vocabulary table quantized to 8 bits too (GatherBlockQuantized), which brings the file from 2.4 GB to 680 MB.

Quality

On Efitra's 103 test texts (English, French, Spanish), compared with the original fp32 model:

  • 99 of 103 identical.
  • No personal detail missed.
  • The 4 differences: 3 extra things hidden (a postcode-like number, a medical word, an amount) and one word ("française") no longer hidden.

Files

File SHA-256
pii_tracer.onnx ab1fd2b4cb2dced8b32ef784d0e07559984fd641566ee12b7e608ee7c45c73d2
pii_tracer.onnx.data d93b74fd12582c084e790c11224e448fc277ffc5a5739df0ab035164d4d59d37
tokenizer.json (unchanged from the original) cae14d1c8dda080f23792355b0692b826bf1f1da3c86ebc1b37548a391cf6526

Tested with ONNX Runtime 1.31 on CPU.

Licence

MIT, as the original. Copyright (c) 2026 Perplexity AI, Inc. See LICENSE. Published by Malaloon Studio.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for malaloon/PII-Tracer-ONNX-Q8

Quantized
(6)
this model