Hugging Face Overview
eternisai/Anonymizer-0.6B: On-Device PII Anonymization
The eternisai/Anonymizer-0.6B is a lightweight, privacy-preserving language model (0.6 billion parameters, Qwen3-based) engineered for surgical anonymization of personal data. Its core function is to detect and replace sensitive information (e.g., names, companies, identifiers, financials) with semantically similar alternatives, ensuring the original query's intent and meaning are preserved.
Key Capabilities & Features
- On-device privacy: Designed to protect user queries by anonymizing PII before data leaves the device.
- Low Latency: Optimized for speed, achieving <200ms TTFT and near-instant completion when quantized, making it suitable for mobile and edge applications.
- Context Preservation: Replaces PII with alternatives that maintain the semantic context required for effective downstream processing.
- Structured Output: Utilizes tool calling to output detected PII and its replacements in a structured JSON format.
- Specific Replacement Rules: Follows detailed rules for replacing various PII types, including names, companies, locations, dates, identifiers, and monetary values, while avoiding public figures or major entities.
Intended Use Cases
- Primary: Integration within the Eternis app for protecting user queries sent to larger LLMs.
- Secondary: Standalone anonymizer for research or other privacy-preserving workflows.
Limitations
- May perform less accurately on nuanced anonymization tasks compared to larger 1.7B and 4B variants.
- Can occasionally miss edge cases involving rare identifiers or subtle contextual PII.
Note: This model requires specific formatting using the tokenizer's chat template with a defined tool schema and a /no_think marker appended to user queries for proper functionality.