Detect and Extract PII from Text Using Hugging Face

This promptfoo example (huggingface/pii) tests whether an LLM's outputs leak personally identifiable information, by scoring them with bigcode/starpii, a model trained to detect PII in source code. The source explicitly flags that starpii was trained for code, not prose, so its suitability for scoring the prose outputs in this example needs to be validated before relying on it. Because the model's card lists no ready-made Inference Provider deployment, you have to obtain access to the gated model yourself and deploy a compatible token-classification endpoint - a Hugging Face Inference Endpoint is one option - that returns the model's entity labels and scores.

Use it as a starting template for scoring LLM outputs against a PII-detection classifier, when you're prepared to validate that classifier's fit for your content type and deploy your own inference endpoint for it. It's not a drop-in PII check: starpii is gated, requiring you to obtain access, has no default hosted deployment, and swapping in a different classifier model requires recalibrating the assertion threshold since the labels and score ranges change.

Input: environment variables HF_STARPII_ENDPOINT, your deployed classifier's URL, HF_TOKEN or HF_API_TOKEN, authorized to call it, and OPENAI_API_KEY for the LLM under test, plus a promptfooconfig.yaml kept consistent with the deployed model's PII labels. Output: standard promptfoo eval results, viewable with promptfoo view, scored against the classifier's entity labels and confidence scores.

• promptfoo eval framework (promptfoo eval, promptfoo view)
• Hugging Face's bigcode/starpii model, deployed as a token-classification Inference Endpoint
• The LLM provider being tested, OpenAI in the example's environment variable

Teams that need a PII-leakage check for LLM outputs and are willing to deploy and validate their own classifier endpoint, rather than teams needing a zero-setup PII scanner.