eternisai/Anonymizer-4B
The eternisai/Anonymizer-4B is a 4 billion parameter language model based on Qwen3, specifically fine-tuned for high-accuracy PII anonymization. It excels at semantically similar replacement of personal data, achieving a 9.55/10 anonymization quality score. This model is designed for critical anonymization tasks in enterprise and research, offering performance comparable to much larger models.
Loading preview...
Model Overview
The eternisai/Anonymizer-4B is a 4 billion parameter language model, part of the Enchanted anonymizer series, developed by eternisai. Built upon the Qwen3-4B architecture, this model is specifically designed for high-accuracy anonymization of Personally Identifiable Information (PII).
Key Capabilities
- High-Accuracy PII Replacement: The model identifies and replaces PII with semantically equivalent alternatives, preserving context while enhancing privacy. It achieves a 9.55/10 score on anonymization quality.
- Efficient Performance: Despite its strong performance, it offers low latency, with Time To First Token (TTFT) under 250ms and full completion under 2 seconds when quantized.
- Structured Output: It generates structured JSON outputs via tool calls, detailing original PII and its anonymized replacements.
Training Details
Anonymizer-4B was trained using Supervised Fine-Tuning (SFT) followed by GRPO (Generative Reinforcement Learning from PPO) with GPT-4.1 acting as the judge. The training dataset comprised approximately 30,000 samples covering various PII replacement and non-replacement scenarios.
Intended Use Cases
- Primary: Integrated as a high-accuracy anonymizer within the Enchanted platform.
- Secondary: Suitable for enterprise and research deployments where top-tier anonymization quality is critical.
Important Usage Notes
- Chat Template Required: The model necessitates the use of
tokenizer.apply_chat_template()with a specific tool schema; raw prompts are not supported. - Special Marker: User queries must include the
/no_thinkmarker for proper PII detection.
Limitations
As the largest model in its series, Anonymizer-4B requires MacBook-class hardware or above for real-time inference and is not optimized for mobile devices as of August 2025.