Teaching LLMs to cosplay as anime protagonists

What it does Chat-Haruhi-Suzumiya is a role-playing chatbot framework that resurrects anime and fictional characters through large language models. It mimics specific personalities, speech patterns, and story context so you can converse with entities like Haruhi Suzumiya or characters from Chinese sitcoms. The project exposes a unified interface over both cloud APIs and locally fine-tuned models, letting you run character simulations against OpenAI, GLM, Baidu, iFlytek, or your own Qwen and ChatGLM2-LoRA checkpoints.

The interesting bit Rather than treating role-play as a prompting parlor trick, the team built an end-to-end production line: script extraction with ASR and voiceprint recognition, a 118K-turn expanded dataset, and fine-tuning pipelines for small models down to Qwen-1.8B. They even spun off psychology research on personality openness, suggesting the project is as much reproducible science as it is fan service.

Key highlights

• Covers 142+ characters (including 95 English roles adapted from RoleLLM) with tone- and plot-aware dialogue
• Backend-agnostic design: swap between OpenAI, ChatGLM, iFlytek, Baidu, and local fine-tuned weights without rewriting chat logic
• Published datasets (54K and 118K turns) and training notebooks for Chinese and English small models
• Bundled toolchain for scraping and cleaning character scripts from raw video and audio
• Apache 2.0 code, though data is CC BY-NC 4.0 and you remain on the hook for original character copyrights

Caveats

• Explicitly a work in progress: version 1.0 is deprecated and the 2.0 refactor is ongoing
• The zero-shot Haruhi-Zero model has migrated to its own repository
• Most cloud demos require third-party API keys; local fine-tuned models need GPU resources

Verdict A solid starting point if you’re building character-driven AI, studying LLM personality alignment, or need Chinese-centric role-play corpora. Look elsewhere if you want a polished consumer app—this is research infrastructure with anime flair.