eternisai/Anonymizer-1.7B: PII Anonymization SLM

The eternisai/Anonymizer-1.7B is a 1.7 billion parameter Small Language Model (SLM) built on the Qwen3 architecture, specifically engineered for robust PII anonymization. It excels at replacing sensitive information with semantically similar alternatives, preserving context while ensuring privacy.

Key Capabilities & Features

• High Accuracy Anonymization: Achieves near-perfect anonymization quality, scoring 9.20/10 against a GPT-4.1 judge (which scores 9.77/10).
• Optimized for Deployment: Balances performance with efficiency, making it suitable for deployment on consumer devices with low latency (e.g., <250ms TTFT, <1s full completion when quantized).
• Advanced Training Methodology: Utilizes Supervised Fine-Tuning followed by Group Relative Policy Optimization (GRPO) with GPT-4.1 as the evaluation judge.
• Structured Output: Generates structured tool calls in JSON format, specifying original PII and its anonymized replacement, facilitating programmatic integration.
• Comprehensive PII Handling: Includes detailed rules for replacing names, companies, projects, locations, dates, identifiers, monetary values, and text snippets.

Ideal Use Cases

• Privacy-Preserving Applications: Primary use as an anonymizer within applications like Enchanted, where sensitive query protection is critical.
• Local Data Processing: Suitable for custom deployments requiring high accuracy PII anonymization directly on local devices.
• Developer Tooling: Provides a robust solution for developers needing to integrate PII masking into their workflows, especially with its structured output and clear usage guidelines.

Important Usage Notes

• Requires specific formatting using tokenizer.apply_chat_template() with a provided tool schema.
• User queries must include the /no_think marker for optimal PII detection.
• The model outputs structured tool calls, which need to be parsed to extract replacements.