h2oai/h2o-llmstudio

H2O LLM Studio - a framework and no-code GUI for fine-tuning LLMs. Documentation: https://docs.h2o.ai/h2o-llmstudio/

H2O LLM Studio – No‑code fine‑tuning for large language models

What it is – A Python‑based framework that ships with a graphical web UI (built on H2O Wave) for training, evaluating, and exporting LLMs. It lets users fine‑tune any Hugging‑Face‑compatible model without writing code, supporting modern techniques such as LoRA, 8‑bit training, DeepSpeed sharding, and the newer DPO/IPO/KTO preference‑based optimizers.

Key capabilities

  • Zero‑code UI – Upload a dataset, pick a model, set hyper‑parameters, launch training, and watch live logs/metrics.
  • Advanced fine‑tuning tricks – LoRA, 8‑bit quantisation, DeepSpeed sharding, and DPO/IPO/KTO (experimental RLHF replacement).
  • Evaluation & comparison – Built‑in metrics, visual experiment comparison, optional Weights & Biases integration.
  • Interactive chat – After training, launch a chat window to test the model instantly.
  • Export – Push the resulting checkpoint to the Hugging Face Hub with a single click or CLI command.
  • CLI mode – All UI actions are reproducible via a YAML‑driven command‑line interface for scripting or headless servers.

Typical workflow

  1. Install (recommended via uv on Python 3.10) and start the Wave server with make llmstudio.
  2. Open http://localhost:10101/ in Chrome.
  3. Create an experiment: select a base LLM, upload data (CSV/JSONL in the format described in the docs), choose a problem type (e.g., causal classification, regression, DPO), and configure hyper‑parameters.
  4. Launch training – the UI shows GPU usage, loss curves, and evaluation scores.
  5. When training finishes, use the built‑in chat to interact with the model or compare it against other runs.
  6. Export the model to Hugging Face or download the checkpoint locally.

Installation highlights

  • System requirements – Ubuntu 16.04+ with an NVIDIA GPU (CUDA 13, driver ≥ 580). 24 GB GPU memory is recommended for larger models.
  • Quick start – make setup creates a uv virtual environment and installs dependencies.
  • Docker alternative – Pull h2oairelease/h2oai-llmstudio-app:latest and run with --runtime=nvidia; the UI is then reachable at the same localhost address.

Running the CLI

uv run python llm_studio/train.py -Y path/to/config.yaml   # single‑GPU
bash distributed_train.sh 4 -Y path/to/config.yaml          # multi‑GPU DDP
uv run python llm_studio/prompt.py -e my_experiment        # chat with a trained model
uv run python llm_studio/publish_to_hugging_face.py -p my_experiment -a <HF_KEY>

Documentation & community

  • Full docs: https://docs.h2o.ai/h2o-llmstudio/
  • Discord for support and discussion.
  • Example notebooks on Kaggle, Colab, and a RunPod cloud template for one‑click GPU instances.

License – Apache 2.0.

Bottom line – H2O LLM Studio provides a beginner‑friendly, yet production‑capable, environment for fine‑tuning state‑of‑the‑art LLMs, handling everything from data ingestion to model export with optional low‑level CLI control for power users.

Related

  • Dispatch
    Hugging Face Guide: Training a LLaMA 2 Chatbot Without Code

    Hugging Face provides a no-code workflow using Spaces, AutoTrain, and ChatUI to allow non-engineers to fine-tune LLaMA 2 and deploy it as a functional chat application.

  • Project
    huggingface/autotrain-advanced

    A no-code tool for training and deploying machine learning models, supporting a wide range of tasks from LLM fine-tuning to image classification.

  • Project
    Lightning-AI/litgpt

    LitGPT is a high-performance, abstraction-free toolkit for pretraining, finetuning, and deploying over 20 popular LLMs at enterprise scale.

  • Dispatch
    Hugging Face Skills: Enabling AI Agents to Fine-Tune LLMs

    Hugging Face has introduced Hugging Face Skills, a tool that allows coding agents like Claude Code to manage the entire LLM fine-tuning lifecycle, from hardware selection to model deployment.

  • Project
    modelscope/ms-swift

    A scalable infrastructure for fine-tuning and deploying 1,000+ text and multimodal large models, supporting the full pipeline from training to quantization and inference.