dylanmurzello/redax-8b

redax-8b: Schema-Driven PII Extraction

redax-8b is an 8 billion parameter language model developed by Dylan Murzello, fine-tuned from Qwen3-8B. Its core function is to accurately identify and extract personally identifying information (PII) from text based on a provided schema. The model outputs a JSON array of {"text": ..., "label": ...} objects, containing the exact character-for-character substrings it identifies.

Key Capabilities

• Schema-Conditioned Span Extraction: The model is designed to operate with a system prompt that defines a schema, specifying labels to find and guard rules for lookalikes that should be ignored.
• Precise Output: It returns exact text spans without rewriting, ensuring fidelity to the original text.
• Robust Negative Handling: A significant portion of its training data includes 'traps' (e.g., clinical values, order numbers) that resemble sensitive data but are not, allowing it to correctly return [] when no PII is found.
• Efficient Deployment: Available in bf16 reference weights and quantized GGUF formats (Q4_K_M for laptops, Q8_0 for near-lossless performance).

Training and Limitations

The model was fine-tuned using full-parameter SFT on 53,141 schema-conditioned examples, combining public corpora (Nemotron-PII, Gretel) with targeted synthetic generation. It was trained for 2 epochs with an effective batch size of 32 and achieved a final eval loss of 0.0185. redax-8b is English-only and is not a compliance tool; it is intended as a component in a larger de-identification system, recommending human oversight or ensemble methods for regulatory compliance (e.g., HIPAA, GDPR).