Hugging Face Launches ML Intern CLI Agent, Lifting GPQA Scores to 32%
Updated
Updated · KDnuggets · Jul 6
Hugging Face Launches ML Intern CLI Agent, Lifting GPQA Scores to 32%
2 articles · Updated · KDnuggets · Jul 6
Summary
Hugging Face released ML Intern, an open-source command-line agent that turns plain-English prompts into machine-learning workflows such as dataset search, script writing, training runs and model publishing.
Built on Hugging Face’s own stack, the tool can search the Hub and arXiv, launch GPU jobs with HF Jobs, log experiments with Trackio, use local or hosted models, and upload outputs back to the Hub.
Hugging Face said the agent improved from about 10% to about 32% on the GPQA scientific-reasoning benchmark in under 10 hours using a small Qwen model, underscoring its iterative workflow rather than one-shot code generation.
The CLI supports an interactive mode with approval checks for risky actions and a headless mode that auto-approves tasks for CI-style automation, with up to 300 turns by default and trace logging for debugging.