Fine-tuning every major LLM into a therapist

What it does EmoLLM is a collection of fine-tuned models and training configurations built for psychological counseling. It wraps general-purpose LLMs like InternLM, Qwen, and LLaMA with domain-specific datasets and XTuner configs to produce conversational mental health assistants. The project also includes evaluation frameworks, pre- and post-training guides, and a newer “CareYou” sub-project that layers RAG, web search, and text-to-speech on top of the base model.

The interesting bit Instead of committing to one foundation model, the authors treat emotional support as a transferable capability, applying QLoRA, LoRA, and full fine-tuning across more than a dozen architectures. The work is backed by an actual ACL Findings paper and a dedicated evaluation benchmark, which is unusual for a community fine-tuning repository.

Key highlights

• Covers a wide model zoo—InternLM2.5, Qwen2, DeepSeek-R1, Baichuan2, GLM-4, Mixtral, and LLaMA3—via XTuner, LLaMA-Factory, and Swift recipes.
• Ships with the multi-turn Chinese counseling dataset CPsyCounD and a DeepSeek-R1 distilled mental health dataset.
• Published evaluation framework (ESC-Eval) and an ACL Findings paper provide empirical backing for the training pipeline.
• CareYou branch adds RAG, web search, and TTS for a more complete assistant stack.

Caveats

• The README and primary datasets are Chinese-first, so developers without reading ability will need translation tools to navigate configs and documentation.
• Much of the value is in curated configs and datasets rather than novel training infrastructure; the heavy lifting is done by existing tools like XTuner.

Verdict Useful if you are building Chinese-language mental health applications or need a reference for domain-specific LLM fine-tuning. Less compelling if you need a single, batteries-included English model with unified documentation.