EmoLLM is a suite of large language models designed for mental health support, covering pre-training, fine-tuning, dataset curation, evaluation, deployment, and RAG. It offers a comprehensive framework for understanding, supporting, and assisting users in mental health contexts, targeting researchers and developers in the AI and mental health fields.

How It Works

EmoLLM leverages instruction fine-tuning on various base LLMs, including InternLM, Qwen, Baichuan, DeepSeek, Mixtral, LLaMA, and GLM series. The project provides fine-tuning configurations for full parameter tuning, LoRA, and QLoRA methods, enabling users to adapt models for specific mental health dialogue tasks. It also includes RAG pipelines and evaluation methodologies for assessing performance in emotion support conversations.

Quick Start & Requirements

• Installation: Clone the repository: git clone https://github.com/SmartFlowAI/EmoLLM.git
• Prerequisites: A100 40G GPU is recommended for InternLM2_7B_chat + qlora fine-tuning with DeepSpeed zero2 optimization. Specific hardware requirements vary by model and fine-tuning method.
• Resources: Detailed guides for quick start, data building, fine-tuning (xtuner, ms-swift, LLaMA-Factory), deployment (LMDeploy), RAG, and evaluation are available within the repository.
• Links: Quick Start Guide, Baby EmoLLM Notebook, Fine-tuning Guide.

Highlighted Details

• Supports a wide range of base LLMs including InternLM2, Qwen2, LLaMA3, Mixtral, and DeepSeek MoE.
• Offers multiple fine-tuning strategies: full parameter, LoRA, and QLoRA.
• Includes dedicated datasets and evaluation frameworks for mental health conversations (e.g., CPsyCoun).
• Provides deployment options via LMDeploy and RAG pipelines.

Maintenance & Community

The project is actively maintained with frequent updates and contributions from a large community. It has received awards and media coverage, indicating strong community engagement and recognition. Links to community discussions or support channels are not explicitly provided in the README.

Licensing & Compatibility

The project is licensed under the MIT License, which permits commercial use and linking with closed-source projects.

Limitations & Caveats

EmoLLM models are intended for emotional support and advice, not professional psychological counseling or therapy. They may produce incorrect, harmful, or offensive output and should not be solely relied upon for critical decisions. Users are advised to exercise caution and discretion.