Harsh221/qwen2.5-1.5b-pii-redactor
Harsh221/qwen2.5-1.5b-pii-redactor is a 1.5 billion parameter Qwen2.5-Instruct model fine-tuned for identifying and redacting Personally Identifiable Information (PII) from text. This model processes raw text and replaces detected PII spans with typed redaction tags, preserving non-PII content. It is specifically optimized for PII detection and text redaction tasks, trained on the ai4privacy/pii-masking-300k dataset.
Loading preview...
Model Overview
Harsh221/qwen2.5-1.5b-pii-redactor is a specialized 1.5 billion parameter model, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct, designed for PII detection and redaction. It transforms raw text by replacing sensitive information with specific redaction tags (e.g., [GIVENNAME1], [EMAIL], [TEL]), while leaving non-PII content untouched. The model was trained using supervised fine-tuning with LoRA on the ai4privacy/pii-masking-300k dataset, processing sequences up to 512 tokens.
Key Capabilities
- PII Redaction: Accurately identifies and replaces various PII types with learned redaction tags.
- Text Generation Redaction System: Operates as a text-to-text model, outputting redacted versions of input text.
- LoRA Fine-tuning: Utilizes LoRA adapters for efficient fine-tuning, resulting in approximately 18.46 million trainable parameters.
- Response-Only Loss: Training focuses on the assistant's redacted response, enhancing redaction behavior.
- System Prompt Adherence: Trained with a specific system instruction to act as a data redaction assistant, which should be used during inference for optimal results.
Good For
- Automated PII Masking: Ideal for applications requiring automated removal or masking of PII in text data.
- Data Anonymization: Useful for preparing datasets or communications where privacy is a concern.
- Developers needing a compact PII redactor: Its 1.5B parameter size makes it suitable for scenarios where larger models might be overkill or resource-intensive.
Limitations
It's important to note that this model is a generative PII redactor and not a formal compliance or privacy guarantee. Its performance on PII recall and precision is not directly measured by eval_loss, and it may miss rare PII types. For high-risk privacy workflows, combining it with deterministic validators and human review is recommended.