Hugging Face's ml-intern Autonomously Trains and Ships Models From a Single Prompt
Hugging Face open-sourced an autonomous agent that reads papers, launches training jobs, and ships models through a single terminal command.
PRO- Hugging Face released ml-intern, an open-source autonomous ML engineering agent, now at 10k+ stars.
- Built on smolagents, it reads papers, prepares datasets, trains models, and launches HF Jobs.
- Runs in interactive or headless mode via a single
uv tool installcommand. - Swappable backends: Claude, GPT, HF Router models, plus local ollama, vllm, lm_studio, llamacpp.
- Includes doom loop detection, approval gates, 170k-token auto-compaction, and MCP server support.
- Sessions auto-upload as private HF datasets viewable in the Agent Trace Viewer.
Hugging Face just released ml-intern, an open-source autonomous agent that behaves like a junior ML engineer sitting inside your terminal. Give it a goal, and the agent researches, writes code, runs experiments, launches jobs on Hugging Face infrastructure, and pushes the final model, streaming every step back as it works. The repo is Apache 2.0, has crossed 10k stars, and installs with a single uv command.
What separates it from a generic coding agent is the plumbing around the LLM, not the model itself. When ml-intern needs a dataset, it searches HF Datasets. When it needs to understand an architecture, it reads the actual paper. When it's ready to train, it launches a real HF Job.
From prompt to shipped model
The tool automates end-to-end post-training workflows for large language models. Built on Hugging Face's smolagents framework, it can autonomously perform literature review, dataset discovery, training script execution, and iterative evaluation, tasks that typically eat significant manual effort from ML researchers and engineers.
The agent runs a continuous loop that mirrors a researcher's workflow. It browses arXiv and Hugging Face Papers, reads methodology sections, and traverses citation graphs to identify relevant datasets and techniques. It searches the Hub for referenced corpora, inspects their quality, and reformats them for training. When local compute isn't available, the agent launches jobs via Hugging Face Jobs.
Getting it running is intentionally boring:
git clone [email protected]:huggingface/ml-intern.git
cd ml-intern
uv sync
uv tool install -e .
ml-intern "fine-tune llama on my dataset"You need an HF_TOKEN for inference and Hub actions, plus a
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
