A chatbot that talks like me: I exported and cleaned my own WeChat chat history, converted it into instruction-tuning data, and LoRA-fine-tuned Qwen2-0.5B-Instruct so the model picks up my real texting style — phrasing, length, and habits included.
Chat logs are messy: stickers, images, fragmented multi-message replies, and mixed contexts. The interesting part of this project is the data pipeline — decrypting and exporting the local WeChat database, filtering message types, merging consecutive messages into coherent question/answer turns, and formatting them as instruction–input–output JSON for supervised fine-tuning.
An end-to-end pipeline: chat-history extraction and cleaning, dataset construction and merging scripts, a LoRA fine-tuning run of Qwen2-0.5B-Instruct (tokenization, prompt templating with chat markup, masked labels), and a small inference app for chatting with the result.
Python · Hugging Face Transformers + PEFT (LoRA) · ModelScope for model download · pandas-style cleaning scripts · the open-source WeChatMsg toolkit for chat export.
【待填写】Repo / demo screenshots.
The 0.5B base keeps training cheap but limits coherence over long conversations; next steps are trying a larger base model, better turn segmentation for group chats, and quantified evaluation of "how much it sounds like me" beyond eyeballing the outputs.