2
Introducing ml-intern, the agent that just automated the post-training team on Hagging Face is an open-source implementation of the real research loop that our ML researchers do every day. You give it a prompt, it researches papers, goes through references, implements ideas in GPU sandboxes, builds in-depth research-backed models for any use case. It’s all built on the Hugging Face ecosystem. It can pull off crazy stuff: – It trained the best model for scientific reasoning. It went through official benchmark paper references. Got OpenScience and NemoTron-CrossThink, added 7 hard-filtered dataset variants from ARC/SciQ/MMLU, and ran 12 SFT runs on Qwen3-1.7B. It increased the score on GPQA by 10% → 32% in less than 10h. Claude Cod’s best: 22.99%. – In healthcare settings he examined available datasets, concluded that they were of very low quality, and wrote a script to generate 1100 synthetic data points from scratch for contingencies, hedging, multilingualism, etc. Then 50x samples were taken for training. Beat Codex by 60% on the health bench. – For competitive math, he wrote a complete GRPO script, started training with A100 GPUs http://hf.co/spacessaw the rewards claimed and then dropped, and continued to run the ablation until it was successful. All fully supported by papers, independently. How does it work? ml-intern makes full use of the HF ecosystem: – searches for papers on arxiv and http://hf.co/papersreads them in their entirety, walks through reference graphs, pulls datasets referenced in procedure sections and on
– Browses the hub, reads recent documentation, inspects datasets and optimizes them before training so it doesn’t waste GPU hours on bad data – Starts training jobs on HF jobs if no local GPU is available, monitors, reads its own eval outputs, diagnoses failures, retrains ML-interns to work in depth and think about how to work It knows what the data should look like and what good models feel like. Releasing it today as a CLI and a web app that you can use from your phone/desktop. CLI:
And the best part? Hugging Face also provided 1k$ GPU resources and Anthropic credits to use it faster.