szc2012/ease-mind

EaseMind is an open source platform that lets anyone train and fine-tune AI language models without needing machine learning experience. It covers the whole process, from downloading a base model to chatting with your finished model, through a web interface rather than the command line. The platform offers two ways to fine-tune a model. A simple mode lets a beginner pick a base model, upload a dataset, choose a quality preset, and start training with one click, with no need to understand any of the underlying parameters. A professional mode exposes nine detailed settings, including LoRA rank, learning rate, and batch size, for users who want full control. There is also a model distillation feature, restricted to admin accounts, where a larger teacher model is used to train a smaller student model. Training progress, including loss curves, is shown live as it happens. Datasets can be uploaded as docx, txt, markdown, csv, json, or jsonl files, or pulled directly from a web page URL, and multiple datasets can be combined in a single training run. Once a model is fine-tuned or distilled, it can be used again as the base for further fine-tuning or distillation. Finished models can be chatted with directly in the browser, with support for multiple turns and formatted text, or connected to outside tools through an OpenAI-compatible API key. The platform can also be connected to DingTalk, Feishu, WeCom, or a generic webhook, turning a trained model into a group chat bot. Administrators can train, distill, and manage models and channels, while regular users are limited to chatting with existing models. The system can run through Docker Compose for an easy setup, or be installed manually with Python. It needs Python 3.10 or newer and works on macOS, Linux, and Windows, with optional GPU acceleration. Under the hood it is built on FastAPI and SQLAlchemy for the backend, PyTorch and Hugging Face Transformers and PEFT for the model training itself, and ModelScope for downloading base models. The frontend is plain HTML, CSS, and JavaScript with no framework. The project is released under the MIT license.