Hugging Face open-source ml-intern, an ML research agent that automatically reads papers, selects data, and runs training

ME News Report, April 22 (UTC+8), according to Beating Monitoring, Hugging Face has open-sourced ml-intern, an ML research agent capable of autonomously completing the entire process of “reading papers, organizing datasets, launching GPU training, evaluating results, and iterative improvement.” The project is built on their own smolagents framework, offering both CLI and web interface options, with the code open on GitHub. The ml-intern toolchain is built around the Hugging Face ecosystem: retrieving papers from arXiv and HF Papers and deep reading along citation chains; browsing datasets on HF Hub, checking quality, reformatting, and reusing for training; when local GPUs are unavailable, calling HF Jobs to start cloud training tasks, automatically reading evaluation outputs, diagnosing failures, and rerunning. By default, it uses Claude Sonnet 4.5 to drive the decision loop, with up to 300 iterations per run, and compresses context exceeding 170k tokens automatically. Hugging Face provided three case studies in their release post. In scientific reasoning tasks, the agent finds datasets like OpenScience and NemoTron-CrossThink from the citation chain of benchmark papers, filters out 7 variants based on difficulty from ARC, SciQ, and MMLU, and runs 12 rounds of SFT on Qwen3-1.7B, increasing GPQA scores from 10% to 32%, taking less than 10 hours. In medical scenarios, the agent judges that existing datasets are insufficient, writes scripts to generate 1,100 synthetic data points, and amplifies the data 50 times for training, surpassing Codex by 60% on HealthBench. In competitive math scenarios, the agent writes GRPO training scripts itself and launches training on A100 via HF Spaces, observing reward collapse and conducting ablation experiments to troubleshoot causes. (Source: BlockBeats)