PII-Tracer-GGUF

This repository provides PII-Tracer, the detector introduced in PII-TRACE.

GGUF packaging of perplexity-ai/PII-Tracer, a PII masking model for conversational data. The model uses a ~600M-parameter bidirectional Qwen3 encoder with a perplexity-ai/pplx-embed-v1-0.6b backbone and two heads:

• Token classification head (1024 → 37): BIOES tags over 9 PII categories (private_person, account_number, private_url, private_date, private_address, private_email, private_phone, other_pii, secret), decoded with a constrained Viterbi decoder.
• Sensitivity head (1024 → 1): conversation-level sensitivity classifier on mean-pooled hidden states.

The backbone is provided as an f16 GGUF for llama.cpp. Both heads remain in fp32 in heads.safetensors. The included Python client applies the heads to per-token hidden states from llama-server, decodes PII spans, and masks text.

Usage

Install a compatible llama-server build and the Python client dependencies, then download the repository:

pip install huggingface_hub torch safetensors transformers
hf download perplexity-ai/PII-Tracer-GGUF --local-dir PII-Tracer-GGUF
cd PII-Tracer-GGUF
llama-server -m pplx-pii-masking-backbone-f16.gguf \
  --host 127.0.0.1 --port 8080 \
  --embeddings --pooling none -c 4096 -b 4096 -ub 4096

Keep the server running. In a second terminal, enter the downloaded model directory and run:

python example_client.py \
  "Hi, I'm Daniel Whitfield, you can reach me at daniels@meridiancap.com or 415-555-0123."

The client prints sensitivity, PII spans, and masked text. It includes the constrained BIOES Viterbi decoder and reads the label list and biases from pii_head_config.json.

Checkpoint layout and inference

The GGUF stores qwen3.attention.causal = false for bidirectional attention. Use --pooling none to obtain one 1024-dimensional hidden state per token. The classification and sensitivity heads are applied separately by the client.

Use llama-server's native /embeddings endpoint for these unnormalized per-token hidden states:

curl -s http://127.0.0.1:8080/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"content":"My email is daniels@meridiancap.com"}'

The client computes token_logits = h @ W_cls.T + b_cls, then decodes the 37 BIOES labels with constrained Viterbi. Sequence sensitivity is sigmoid(mean_pool(h) @ W_sen.T + b_sen). Tokenization and character offsets use the supplied tokenizer files.

Limitations

• Inputs must contain between 1 and 4096 tokens. Chunk longer documents before calling the example client; it sends the entire supplied text to the server while its local offset tokenizer truncates at 4096 tokens.
• Bidirectional attention requires the whole input in one micro-batch. Keep -b and -ub at 4096 for the supported input window.
• /v1/embeddings does not support --pooling none; use /embeddings.
• Conversion precision and the inference backend can affect numerical outputs. Evaluate detection quality and sensitivity thresholds on your target data; false positives and missed PII are possible.

Files

License

MIT, matching the base model. See LICENSE.

Citation

If you use or reference this work, please cite:

@article{zhang2026piitrace,
  title = {{PII-TRACE}: A Benchmark for Context-Aware {PII} Detection in Multi-Turn {LLM} Conversations},
  author = {Zhang, Kaiyuan and Wang, Chuan and Zhong, Joey and Fryzel, Paul and Polley, Kyle and Ma, Jerry and Li, Ninghui},
  journal = {arXiv preprint arXiv:2609.22200},
  year = {2026},
  url = {https://arxiv.org/abs/2609.22200}
}