h2o-llmstudio: a no-code GUI and framework for fine-tuning large language models with support for memory-efficient training

What it solves

H2O LLM Studio is designed to make the fine-tuning of large language models (LLMs) accessible to users regardless of their coding experience. It removes the technical barriers to customizing LLMs by providing a no-code graphical user interface (GUI) and a flexible framework for managing experiments.

How it works

The project provides a GUI and a command-line interface (CLI) to manage the fine-tuning process. Users can upload datasets, configure hyperparameters, and use advanced techniques like Low-Rank Adaptation (LoRA) and 8-bit training to reduce memory requirements. It supports various optimization techniques including DPO, IPO, and KTO for preference optimization, as well as Causal Regression and Classification modeling. For larger models, it integrates DeepSpeed for sharded training across multiple GPUs.

Who it’s for

It is intended for developers and AI researchers who want to fine-tune LLMs without writing code, as well as those who prefer a CLI-based workflow for automated experiments.

Highlights

• No-Code GUI: A dedicated interface for configuring and launching fine-tuning experiments.
• Memory Efficiency: Support for LoRA and 8-bit training to lower GPU memory footprints.
• Advanced Optimization: Includes DPO, IPO, and KTO as alternatives to RLHF.
• Experiment Tracking: Visual tools to track, compare, and integrate with Weights & Biases (W&B).
• Model Export: Direct export capabilities to the Hugging Face Hub.
• Multi-GPU Support: Integration with DeepSpeed for training larger models on multiple GPUs.