huggingface/ml-intern

What it solves

ML Intern is an autonomous agent designed to handle the end-to-end process of researching, writing, and shipping high-quality machine learning code. It removes the manual effort of navigating documentation, searching for papers, and managing cloud compute for ML development.

How it works

The project uses an agentic loop (up to 300 iterations) that integrates with various LLMs via LiteLLM. It employs a ToolRouter to access a wide array of capabilities, including Hugging Face documentation, research papers, datasets, and GitHub code search. The agent can operate in two environments: a local filesystem for editing files in a checkout, or a Hugging Face Space sandbox for remote testing and GPU hardware requests.

Who it’s for

It is built for ML engineers and researchers who want to automate the implementation of ML workflows, such as fine-tuning models or testing training scripts, while leveraging the Hugging Face ecosystem.

Highlights

• Deep Ecosystem Integration: Direct access to HF docs, papers, datasets, and cloud compute.
• Flexible Model Support: Works with hosted inference providers or local models via Ollama, vLLM, LM Studio, and llama.cpp.
• Sandbox Execution: Ability to create and test code in private HF Space sandboxes with GPU support.
• Traceability: Automatically uploads session traces to private Hugging Face datasets for easy browsing and tool-call analysis.
• Doom Loop Detection: Built-in mechanism to detect repeated tool patterns and inject corrective prompts to prevent the agent from getting stuck.