cagrigungor/pii-guard-turkish-270m
The cagrigungor/pii-guard-turkish-270m is a 270 million parameter causal language model, based on the google/gemma-3-270m-it architecture, developed by Hasan Çağrı Güngör. Designed for on-device (mobile) use, this model specializes in instruction-conditional masking of 53 types of Personal Identifiable Information (PII) in Turkish text. It is optimized for data minimization tasks such as pre-LLM data masking, log cleaning, and GDPR/KVKK compliance, offering flexible masking policies based on user instructions.
Loading preview...
Overview
The cagrigungor/pii-guard-turkish-270m is a compact, 270 million parameter model built on the google/gemma-3-270m-it architecture, specifically designed for on-device PII masking in Turkish. Developed by Hasan Çağrı Güngör, this model excels at identifying and masking 53 distinct PII labels within Turkish text, offering instruction-conditional masking capabilities. This means the same text can be masked differently based on the provided policy, making it highly adaptable for various data privacy requirements.
Key Capabilities
- Instruction-Conditional Masking: The model interprets instructions to apply specific masking policies (e.g., full masking, whitelist, blacklist, category-based, out-of-scope). This allows for dynamic control over which PII elements are masked.
- Extensive PII Recognition: It recognizes 53 PII labels across nine categories, including identity, financial, communication, special categories (KVKK md. 6), authentication, device/network, vehicle, employment, and demographic information.
- Turkish-Specific Handling: Features robust handling of Turkish linguistic nuances such as PII with suffixes (e.g., "Ayşe Yılmaz'ın"), PII written out in words, and PII in all caps. It also demonstrates resistance to false positives like project names or generic text containing PII keywords.
- Performance: Achieves a 0.740 exact match score on a 1,000-example, independently curated Turkish PII masking benchmark, with a schema-neutral score of 0.770.
- On-Device Optimization: Designed for mobile use, with a GGUF version available for deployment on phones.
Use Cases
This model is ideal for developers needing to implement a masking layer before sending data to LLMs, for log cleaning, and for ensuring compliance with data protection regulations like KVKK/GDPR through data minimization. Its ability to handle complex Turkish text and conditional masking makes it a powerful tool for privacy-preserving applications.