Hugging Face's ml-intern Runs Your Whole Post-Training Pipeline for You

Hugging Face shipped an autonomous agent called ml-intern that does the unsexy parts of ML research. It searches arXiv and HF Papers, reads citations and methodology sections, finds and reformats datasets from the Hub, launches SFT or GRPO jobs on Hugging Face Jobs, watches the runs, debugs failures, and iterates. Built by Aymeric Roucher (the smolagents author), Aksel Joonas Reedi, and Yoan Di Cosmo. Open source, with a CLI and a Spaces UI.

The meta-loop here is the point. Anthropic, OpenAI, and Google are competing on agents that automate end-user knowledge work. ml-intern automates the work that produces those agents in the first place. Post-training has been the rate limiter for every team trying to build a vertical agent. If you can hand the SFT-eval-debug loop to a competent intern that knows how to read a paper, your iteration speed jumps.

This is also the clearest example yet of Hugging Face owning its own stack: their hub, their datasets, their compute, their orchestration. Pay the bill, get a research assistant that runs all night. Source: https://www.producthunt.com/products/ml-intern