promptrails/piimask-qwen2.5-0.5b
promptrails/piimask-qwen2.5-0.5b is a compact 0.5 billion parameter Qwen2.5-based model developed by promptrails, specifically designed for detecting Personally Identifiable Information (PII) in short, informal English and Turkish texts. It operates fully locally, ensuring data privacy by not sending information off-device. This model excels at identifying nine PII entity types, making it suitable for pre-processing chat messages, support tickets, and emails before further LLM processing or logging.
Loading preview...
Overview
promptrails/piimask-qwen2.5-0.5b is a specialized 0.5 billion parameter model built on the Qwen2.5 architecture, focused on local-first PII detection in English and Turkish. It processes short, informal texts like chat messages, support tickets, and emails, returning a JSON list of identified entities. The model supports nine PII types: PERSON, EMAIL, PHONE, NATIONAL_ID, IBAN, CREDIT_CARD, ADDRESS, DATE_OF_BIRTH, and IP_ADDRESS.
Key Capabilities & Features
- Bilingual PII Detection: Optimized for both English and Turkish short, informal texts.
- Local Operation: Ensures data privacy as all processing occurs on-device without external API calls.
- Specific Entity Types: Identifies a focused set of nine common PII categories.
- Enhanced Turkish Address Detection: Includes a deterministic grammar to significantly improve recall for street-first Turkish addresses, a common challenge for LLMs.
- Strong Performance: Achieves a micro-F1 score of 0.81 on an independent bilingual benchmark, outperforming Microsoft Presidio (0.51) and ai4privacy DeBERTa (0.65) for its target domain.
Intended Use Cases
- PII Scrubbing: Ideal for pre-processing input before sending to third-party LLMs.
- Developer Log Sanitization: Helps cut PII from developer logs for privacy compliance.
- Privacy-Conscious Utilities: Suitable for local applications requiring PII removal.
Limitations
While effective, piimask-qwen2.5-0.5b is intended as one layer in a defense-in-depth strategy, not a sole compliance-grade safeguard. It may miss a meaningful fraction of PII, particularly names and free-text entities in long or formal documents. Users should pair structured entity outputs (EMAIL, IP, IBAN, etc.) with deterministic validation passes for maximum precision.