IAAR-Shanghai/MemPrivacy-4B-SFT
MemPrivacy-4B-SFT is a lightweight, privacy-preserving model developed by IAAR-Shanghai, fine-tuned from the Qwen3-4B base model. It specializes in high-precision identification and classification of privacy-sensitive information within conversational text, using a four-level privacy taxonomy. Designed for edge-cloud agents, it replaces sensitive data with semantically structured placeholders to preserve utility while ensuring local privacy protection. This model is optimized for resource-constrained environments and excels at structured privacy information extraction.
Loading preview...
Overview
MemPrivacy-4B-SFT is a specialized, privacy-preserving language model developed by IAAR-Shanghai, built upon the Qwen3-4B base architecture. It functions as the core local extraction engine within the MemPrivacy framework, designed to manage personalized memory in edge-cloud agent systems. The model's primary innovation lies in its ability to accurately identify and categorize privacy-sensitive spans in user dialogues, replacing them with type-aware placeholders (e.g., <Email_1>) before data transmission to the cloud. This approach ensures that task-relevant semantics are preserved, unlike aggressive masking methods.
Key Features & Capabilities
- High-Precision Privacy Extraction: Achieves superior performance in identifying privacy information, outperforming general-purpose reasoning models. It uses a sophisticated four-level privacy taxonomy (PL1-PL4) to classify content based on identifiability, harm, and exploitability.
- Semantic Utility Preservation: By using structured, typed placeholders, the model maintains relational and semantic cues necessary for effective memory formation and personalization by cloud agents.
- Edge-Optimized Efficiency: Engineered for resource-constrained local deployment, it delivers high accuracy with significantly reduced inference latency compared to larger general-purpose LLMs.
Use Cases
- Privacy-Preserving AI Agents: Ideal for applications requiring local processing of sensitive user data before cloud interaction.
- Structured Data Extraction: Capable of extracting and categorizing privacy information into a defined JSON schema, facilitating downstream privacy management and compliance.
- Fine-Grained Privacy Control: Enables user-configurable protection policies based on its detailed privacy taxonomy.