Skip to main content

Quick start: fine-tune your first model

This guide walks you through the basic flow of using H2O Enterprise LLM Studio to fine-tune a large language model (LLM) on your own data — from dataset upload all the way to deployment.

You can get started in just a few steps:

Step 1: Upload or select a dataset​

Every experiment begins with data. From the Datasets tab in your project, click New Dataset to upload your own file (CSV, JSON, or Parquet), or select one of the built-in demo datasets.

If you’re just exploring, the pre-loaded datasets are a great way to try out different types of problems like text generation or classification.

Step 2: (Optional) Generate synthetic data​

If your dataset is missing labels or you want to create a brand new dataset, use the Data Generation tab. You can either:

  • Generate rows (create new synthetic examples using a prompt), or
  • Generate columns (label or annotate existing data)

You’ll configure a prompt and select a large model like GPT-4 or Claude to do the generation work. The output becomes a new dataset ready for training.

Step 3: Create and configure an experiment​

From the Experiments tab, click New Experiment to start training a model.

You’ll choose:

  • A dataset (either one you uploaded or generated)
  • A backbone model to fine-tune (like Falcon, Mistral, or Danube)
  • Hyperparameters — or just leave them at their defaults for now

Advanced users can configure things like learning rate, number of epochs, and LoRA settings.

Step 4: Use AutoML or Ask KGM to improve performance​

To guide your experiment, you have two interactive helpers:

  • AutoML: Before you run your experiment for the first time, select AutoML to automatically test several promising configurations sequentially based on the advice of an AI assistant.
  • Ask KGM: Instead of AutoML, you can use Ask KGM (an AI assistant) to recommend a single new backbone or set of hyperparameters at a time.

Both help you quickly find the best settings and models for your data — great for iterating or if you’re unsure how to tune.

Step 5: Deploy your model​

Once you're happy with the model’s performance, go to the Deployments tab and click New Deployment.

You can:

  • Deploy a model you trained in H2O Enterprise LLM Studio
  • Or deploy a publicly available model from Hugging Face

Once deployed, you’ll get:

  • An API endpoint for integration into downstream apps
  • A chat interface to test your model directly in the UI

That’s it! You’ve fine-tuned and deployed your first LLM in just a few clicks. For deeper customization or advanced workflows, explore the other sections in the docs.

Feedback