Skip to main content
Version: v0.24.x

System overview

Overview​

H2O Cascade is organized around a small set of core components that work together to take you from raw data to a deployed, production-ready model. This page describes how those pieces fit together at a high level. For step-by-step usage of any individual component, see its dedicated page.

Core components​

  • Projects — The top-level workspace boundary. Every dataset, data generation job, experiment, AutoML run, and deployment belongs to a project, and switching projects filters everything else in the UI to that project's resources.
  • Datasets — Training and validation data, either uploaded directly or produced by a Data Generation job. Each dataset carries a Problem Type that determines which experiment settings are available for it.
  • Data Generation — Uses a large, hosted LLM (for example, GPT-4, Claude, or Gemini) to create synthetic rows or label/annotate existing data, producing a new dataset that's ready for fine-tuning.
  • Experiments — Training runs. An experiment trains a backbone model for the dataset's problem type, such as text generation, text classification, image classification, or object detection, and reports evaluation metrics on a validation set. For LLM problem types, the backbone is a model such as Meta Llama, Qwen, or Google Gemma, fine-tuned with LoRA, QLoRA, or full fine-tuning.
  • AutoML — Iterates on Experiments automatically, adjusting hyperparameters and backbone selection across a series of runs to find the best-performing configuration. Ask KGM offers the same kind of recommendation one experiment at a time, from the experiments list or an experiment's detail view.
  • Deployments — Serves a trained experiment (or a public Hugging Face model) behind a REST endpoint and an in-UI chat interface for testing.
  • REST API — Every action available in the UI — creating projects, uploading datasets, running data generation, launching experiments and AutoML runs, and managing deployments — is also available programmatically.

How the pieces connect​

A typical path through the platform touches these components in order: a dataset is uploaded or produced by Data Generation, an experiment fine-tunes a backbone model against that dataset (optionally guided by AutoML or Ask KGM), and the resulting model is served through a Deployment. See Experiment (model) flow for the step-by-step walkthrough of this path.


Feedback