ML-intern-lab

Models, datasets and demos built by an intern that never sleeps.

What is this?

Everything in this organization was researched, trained, evaluated and pushed by ML Intern, the open-source agent from Hugging Face that autonomously researches, writes and ships ML code using the Hugging Face ecosystem, with access to docs, papers, datasets and cloud compute.

The workflow for each repo:

1. Describe the idea in a conversation.
2. The intern researches. It searches the Hub, GitHub, papers and the web for the right models, datasets and tools.
3. Approve a budget. Once approved, the system enforces it. The agent can't change its own budget.
4. The intern ships. It creates the repo, generates data, launches training on HF Jobs, tracks the run on a Trackio dashboard, pushes the weights, writes the report and deploys a demo Space.

What you see here is the output of that loop, unedited.

What's in here

Each model card links to the session trace, the Trackio dashboard and the approved budget for that run.

House rules

• Every run has a trace. The full session log is stored as a Claude Code JSONL dataset and can be opened in the HF Agent Trace Viewer.
• Budget-capped. Nothing here cost more than the amount approved before the run started.
• Evals as measured. Runs that didn't beat the baseline are kept and reported as-is.
• Apache 2.0 or whatever the base model's license is.

About

This is not an official Hugging Face or ml-intern team org. The goal is to find out how far an autonomous ML agent can go on small models and small budgets.

Built on: