| license: mit | |
| base_model: | |
| - perplexity-ai/PII-Tracer | |
| pipeline_tag: token-classification | |
| library_name: onnx | |
| language: | |
| - en | |
| - multilingual | |
| tags: | |
| - pii | |
| - privacy | |
| - token-classification | |
| - webgpu | |
| - onnx | |
| # PII-Tracer WebGPU weights | |
| ONNX derivative weights of [`perplexity-ai/PII-Tracer`](https://huggingface.co/perplexity-ai/PII-Tracer) | |
| for browser inference on WebGPU or CPU. `catalog.json` lists the available | |
| variants, the graph's inputs and outputs, and the label set. See `LICENSES.md` | |
| for provenance and license details. | |
| The graph takes `input_ids` alone — the bidirectional attention mask the | |
| encoder-only backbone needs is built inside it — and returns per-token BIOES | |
| logits over nine PII types plus a sequence-level sensitivity score. Weights sit | |
| beside each graph in a `.onnx_data` file that must be fetched with it. | |
| Decoding needs the constrained Viterbi step from the original model card; a | |
| per-token argmax admits sequences BIOES cannot produce. | |