Hugging Face Open Sources ml-intern, an ML Research Agent for Automated Paper Reading, Data Selection, and Training

According to monitoring by Dongcha Beating, Hugging Face has open-sourced ml-intern, an ML research agent capable of autonomously completing the entire process of “reading papers, organizing datasets, initiating GPU training, evaluating results, and iterating improvements.” The project is built on their own smolagents framework and offers both CLI and web-based access, with the code available on GitHub. The toolchain of ml-intern is constructed around the Hugging Face ecosystem: it retrieves papers from arXiv and HF Papers and conducts deep reading along citation chains; it browses datasets on HF Hub, checks quality, reformats them, and then inputs them for training; when there is no local GPU, it can call HF Jobs to initiate cloud training tasks, automatically reading evaluation outputs, diagnosing failure reasons, and rerunning after training is completed. By default, it uses Claude Sonnet 4.5 to drive the decision loop, with a maximum of 300 iterations per run and automatic compression of context exceeding 170k tokens. Hugging Face provided three case studies in their release post. In a scientific reasoning task, the agent identified the OpenScience and NemoTron-CrossThink datasets from the citation chain of a benchmark paper, filtered out seven variants from ARC, SciQ, and MMLU based on difficulty, and ran 12 rounds of SFT on Qwen3-1.7B, increasing the GPQA score from 10% to 32% in under 10 hours. In a medical scenario, the agent determined that the quality of existing datasets was insufficient and autonomously wrote a script to generate 1,100 synthetic data points, expanding the dataset by 50 times for training, surpassing Codex by over 60% on HealthBench. In a competitive math scenario, the agent independently wrote a GRPO training script and initiated training on an A100 through HF Spaces, observing reward collapse and conducting ablation experiments to investigate the cause.