Skip to main content
A Harbor task is an instruction, sandbox environment, and test script. A dataset is a collection of tasks. Datasets sometimes define custom metrics that aggregate task rewards. There are three ways to use datasets:
  1. Local datasets: run a local directory of tasks or dataset.toml manifest
  2. Published datasets: run a dataset from Harbor Hub
  3. Git repository datasets: run a dataset from a Git repository

​
Local datasets

Run a local dataset with --path or -p:
The path can be a directory of tasks or a dataset.toml manifest file.

​
Published datasets

Datasets can be published and shared with members of your organization or publicly on Harbor Hub. Run a published dataset with --dataset or -d:
To learn how to create and publish a dataset, see Publish a dataset.

​
Git repository datasets

Run a dataset directly from any Git repository with --repo:
Supports GitHub, GitLab, and Hugging Face URLs, with optional ref pinning (@v1.0, @main). See Git repos for the full guide.