A chatbot that learned its personality from Bilibili comments

What it does

This project takes Alibaba’s Qwen1.5-32B-Chat and fine-tunes it on Bilibili user comments using Apple’s MLX LoRA pipeline, producing a local chatbot that responds with the tone of the platform’s comment sections. It pairs text generation with GPT-SoVITS for optional voice output, using either the author’s own voice model or a community-trained Paimon voice. A compress_model.py utility quantizes the merged model to accelerate local inference.

The interesting bit

Instead of aiming for a helpful assistant, the project treats Bilibili comments as a personality dataset, turning a 32B-parameter base model into a character engine. The included voice sample suggests it is not afraid to tell you when a question is underwhelming.

Key highlights

• Fine-tunes a 32B model locally via Apple’s MLX framework
• Optional voice synthesis through GPT-SoVITS, including a third-party Paimon voice model
• Bundled quantization script to compress the merged adapter model
• Supports both interactive chat and batch dialogue generation from questions.txt
• Runs fully offline after initial model setup

Caveats

• The project is largely glue code connecting MLX fine-tuning with GPT-SoVITS inference
• Voice setup is manual: it requires a separate GPT-SoVITS environment and running two API servers on different ports
• The README only confirms macOS testing, and a 32B model will need substantial memory

Verdict

A fun experiment for developers interested in Bilibili comment culture as an LLM personality layer, or anyone wanting a local Chinese-speaking character bot with voice output. Not the right choice if you need a polished, general-purpose assistant.