H2O LLM Studio is a framework and graphical interface for fine-tuning large language models without writing code. You upload a dataset, configure an experiment through a GUI built on H2O Wave, train, and then compare runs, chat with the result, and publish it to the Hugging Face Hub.
Training supports Low-Rank Adaptation (LoRA), 8-bit training for a smaller memory footprint, and preference optimization with DPO, IPO, and KTO (these replaced the earlier RLHF mode). Problem types go beyond chat tuning to causal classification (binary and multiclass) and causal regression with an LLM backbone. DeepSpeed provides sharded training across multiple GPUs, and a CLI runs the same experiments from a YAML configuration, including multi-GPU data-parallel runs.
It runs on Linux machines with a recent NVIDIA GPU (24 GB of GPU memory or more is suggested for larger models), in Docker, or on cloud instances such as RunPod, Kaggle, and Colab. Experiment tracking can be sent to Weights & Biases. The project notes that backward compatibility is not guaranteed during rapid development.
Features
- No-code GUI: configure, run, monitor, and compare fine-tuning experiments in a browser
- Hyperparameter control: a wide range of training settings for any supported LLM
- Efficient fine-tuning: LoRA and 8-bit training to reduce memory use
- Preference optimization: DPO, IPO, and KTO, including KTO pair loss for simple preference data
- Classification and regression: causal classification and causal regression problem types built on LLMs
- Multi-GPU training: DeepSpeed sharded training and DDP mode
- Evaluation: metrics that judge generated answers, plus visual performance tracking
- Interactive chat: talk to a trained model from the GUI or CLI for quick feedback
- Hugging Face export: publish fine-tuned models to the Hugging Face Hub
- Experiment tracking: optional Weights & Biases integration
