H2O LLM Studio

A self-hosted LLM fine-tuning tool with a no-code browser interface, memory-saving training methods and support for NVIDIA GPUs on Ubuntu or Docker.

H2O LLM Studio helps teams adapt language models to their own datasets through a browser interface, without writing training code. It's aimed at organizations building models for specific tasks.

Model fine-tuning: Train models for instruction and chat tasks, or adapt small language models for classification and regression.
Memory-saving training: LoRA and 8-bit training reduce the GPU memory needed to fine-tune models.
Evaluation and comparison: Score generated answers, compare experiments visually and chat with a trained model to test its responses; Weights & Biases integration adds experiment tracking.
Preference alignment: DPO, IPO and KTO training methods help adapt responses to preference data.
Model sharing: Export trained models to Hugging Face Hub for others to use.

The project is open source under Apache 2.0. It runs on Ubuntu with a recent NVIDIA GPU and supports Docker and multiple GPUs. For larger models, the project recommends at least 24 GB of GPU memory. Training can run on your own hardware or a cloud GPU machine such as RunPod. A command-line interface also supports work outside the browser.