Hugging Face Automates LLM Post-Training With Open-Source Agent
Hugging Face has released ml-intern, an open-source AI agent that automates the full post-training workflow for large language models. Built on the company's smolagents framework, it runs as a continuous loop: browsing arXiv and Hugging Face Papers, sourcing and reformatting datasets from the Hub, launching compute jobs, and retraining until benchmark scores improve.
MarkTechPost reports the agent was evaluated against PostTrainBench, a benchmark developed by researchers at the University of Tubingen and the Max Planck Institute. The benchmark constrains the agent to a single H100 GPU and a strict 10-hour window. Experiment tracking runs through Trackio, a Hub-native tool positioned as an open-source alternative to Weights and Biases.
The operating consequence is direct: tasks that required a dedicated ML engineer now run autonomously. Labs with smaller headcounts can iterate on post-training at a pace previously reserved for well-staffed research teams. For French AI operators building on open infrastructure, the question shifts from staffing the workflow to owning the evaluation criteria that agent optimizes toward.