AIME: Fine-Tuning Large Models with Personal iMessage History to Create a Custom AI Chat Style

The AIME project extracts macOS iMessage chat history, processes and formats the data with Claude, and fine-tunes the Gemma 4 model using QLoRA technology to create a personalized AI that mimics the user's personal chat style. This project is an end-to-end solution covering the entire workflow of data extraction, processing, and training.

With the improvement of large model capabilities, users expect AI to have personal styles (slang, emojis, tone differences, etc.). Traditional general AI assistants have standardized tones and lack personality; the AIME project provides a solution to extract style features from personal chat records and fine-tune models.

Data Pipeline: Extract macOS iMessage database records → Claude splits conversations → Relationship classification (partner/friend/family, etc.) → Format conversion to dialogue JSONL → Data review and cleaning → QLoRA fine-tuning of the Gemma 4 31B model. Technical Highlights: Native macOS integration (requires disk access permission), Claude-driven intelligent processing, breakpoint resume mechanism, relationship-aware training data (annotated with relationship types).

• Local-first architecture: Data is only sent to AWS when Claude is called, and users control their data;
• Automatic PII desensitization: Identifies and desensitizes sensitive information such as emails and addresses during the cleaning phase;
• Editable relationship mapping: Users can manually review and modify contacts/relationship_map.json.

• Personal AI assistant: Reply to messages and write emails in the user's style;
• Style transfer research: Provide data for studies on language style and social relationships;
• Creative writing assistance: Help writers create dialogues that fit character settings;
• Social skill training: Gain insights by analyzing expression methods in different relationships.

• Platform limitations: Only supports macOS and iMessage;
• Computational cost: High cloud GPU training fees;
• Data quality dependency: Model performance is affected by the quantity and diversity of chat records;
• Privacy trade-off: Requires trust in AWS's privacy policy (Claude processing relies on cloud services).

• Support for more messaging platforms (WhatsApp, WeChat, etc.);
• Incremental training: Regularly update the model with new messages;
• Hybrid personality model: Combine multiple personal styles;
• Fine-grained style control: Dimensions such as emotional state and formality level.

The AIME project demonstrates a feasible prototype for extracting style features from personal data. Although it has platform limitations and privacy considerations, it provides a workable solution for the vision of "AI being like oneself". With the improvement of local small model capabilities and fine-tuning technology, personalized AI solutions will become more popular.