ml-intern

// Overview

ml-intern is an open-source agent that does machine-learning engineering work end to end rather than answering questions about it. Given a goal, it researches — reading documentation and papers, looking for datasets and existing models on the Hub — then writes the code, runs it against local or cloud compute, reads the results and iterates. The output is a trained model and the code that produced it, not a suggestion.

It runs as a terminal application, either interactively as a chat or headlessly against a single prompt, which is what makes it usable for long jobs you leave running. Authentication is a Hugging Face token plus a GitHub token; models can come from a hosted provider or from any OpenAI-compatible endpoint, so Ollama, vLLM, LM Studio and llama.cpp all work as the driving model.

Practical guardrails come from the agent runtime rather than the prompt: tools can execute in a sandbox or against the local filesystem, sensitive operations sit behind approval gates, and a doom-loop detector interrupts the agent when it starts repeating the same tool pattern. Sessions can be auto-uploaded to a private Hugging Face dataset so a long run is reviewable afterwards, and MCP servers can be attached to give the agent extra tools.

// What it does

• Runs the full loop — literature and docs research, dataset discovery, training code, execution, evaluation — as one agent
• Interactive chat mode and headless single-prompt execution for unattended runs
• Driven by a hosted model or any OpenAI-compatible local endpoint (Ollama, vLLM, LM Studio, llama.cpp)
• Sandbox or local-filesystem tool runtimes, with approval gates on sensitive operations
• Doom-loop detector that breaks the agent out of repeated tool patterns
• Optional session upload to a private Hugging Face dataset, and one-way notification gateways such as Slack
• MCP server integration for additional tools

// Getting started

The project is installed from a checkout with uv, then exposed as the `ml-intern` command.

git clone git@github.com:huggingface/ml-intern.git
cd ml-intern
uv sync
uv tool install -e .

HF_TOKEN=<your-hugging-face-token>
GITHUB_TOKEN=<github-personal-access-token>

ml-intern

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Hand off a post-training experiment — pick a base model, find data, fine-tune, evaluate — and read the results later
• Sweep a family of training configurations overnight in headless mode instead of babysitting a notebook
• Turn a paper you want to reproduce into a running training script grounded in the Hugging Face ecosystem
• Run the whole loop against a local model endpoint when the data cannot leave your machine

// How ml-intern compares

ml-intern alongside other open-source autonomous coding agents tools AI/TLDR tracks, ranked by GitHub stars.