cagrigungor/pii-guard-turkish-0.6b
The cagrigungor/pii-guard-turkish-0.6b is a 0.6 billion parameter instruction-conditioned model developed by Hasan Çağrı Güngör, specifically designed for masking Personal Identifiable Information (PII) in Turkish text. Utilizing a classic attention architecture, it is optimized for efficient CPU deployment via ONNX export. This model excels at data minimization for compliance (KVKK/GDPR) and log cleaning, recognizing 53 PII labels and adapting its masking behavior based on provided instructions.
Loading preview...
PII Guard Turkish 0.6B: Instruction-Conditioned PII Masking
The cagrigungor/pii-guard-turkish-0.6b is a 0.6 billion parameter model developed by Hasan Çağrı Güngör, specifically engineered for masking Personal Identifiable Information (PII) in Turkish text. It stands out due to its instruction-conditioned masking capabilities, allowing users to define masking policies (e.g., full masking, whitelist, blacklist) through natural language instructions. Built on a classic attention architecture, it offers clean export to ONNX, making it a practical choice for CPU-based deployments.
Key Capabilities & Features
- Instruction-Conditioned Masking: Adapts masking behavior based on user-provided instructions, supporting various policies like full masking, whitelisting, blacklisting, and category-specific masking.
- Extensive PII Recognition: Identifies and masks 53 distinct 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: Demonstrates robust behavior with Turkish linguistic nuances, such as handling PII with suffixes (e.g., "Ayşe Yılmaz'ın"), PII written out in words (e.g., "tc kimlik numaram yedi bir dokuz..."), and PII in all caps.
- High Accuracy: Achieves a full match score of 0.828 on a 1,000-example, manually curated, independent Turkish PII masking benchmark, outperforming other models in its class.
- Efficient Deployment: Its classic attention architecture facilitates clean ONNX export, ideal for CPU inference.
Use Cases & Limitations
This model is primarily designed as a masking layer before sending data to LLMs, for log cleaning, and for KVKK/GDPR data minimization. It is a valuable tool for ensuring data privacy and compliance in Turkish-language applications.
Limitations to consider:
- Synthetic Data Evaluation: Performance metrics are based on synthetic data; real-world performance may vary and should be validated with your own data.
- Schema Adherence: May occasionally generate out-of-schema labels; a whitelist layer is recommended in production.
- Specific Weaknesses: Exhibits lower scores on blacklist instructions and long, verbally written number sequences.
- Turkish Only: Not tested or optimized for other languages.
- Decision Support: Not a substitute for human oversight in critical compliance workflows.