Hugging Face Launches ml-intern to Automate LLM Post-Training

Hugging Face has officially released ml-intern, an open-source AI agent designed to autonomously manage the end-to-end post-training workflow for large language models. Built on the company’s smolagents framework, the tool replicates the typical research cycle by performing literature reviews, discovering datasets, executing training scripts, and conducting iterative evaluations. By automating these complex tasks, ml-intern aims to reduce the significant manual effort traditionally required by machine learning researchers and engineers.

Automating the Research Loop

The agent functions as a continuous, autonomous loop that mirrors the behavior of an ML researcher. It begins by browsing arXiv and Hugging Face Papers to analyze methodology sections and traverse citation graphs, allowing it to identify relevant techniques and datasets. Once it locates datasets on the Hugging Face Hub, the agent inspects their quality and reformats them for specific training requirements. If local compute resources are unavailable, the agent can autonomously launch jobs via Hugging Face Jobs.
Throughout the process, the agent monitors performance using Trackio, a Hub-native experiment tracker designed as an open-source alternative to Weights & Biases. After each training run, ml-intern reads evaluation outputs to diagnose potential failures, such as reward collapse in reinforcement learning from human feedback (RLHF) pipelines, and continues the cycle until benchmark performance improves.

Performance and Benchmarking

To test its capabilities, the agent was evaluated against PostTrainBench, a benchmark developed by researchers at the University of Tübingen and the Max Planck Institute. This benchmark measures an agent's ability to post-train a base model within a strict 10-hour window on a single H100 GPU. In official demonstrations, ml-intern took the Qwen3-1.7B base model—which has a baseline score of roughly 10% on the GPQA benchmark—and improved it to 32% in under 10 hours. Notably, the agent reached a 27.5% score in just over three hours.
These results highlight the agent's data efficiency, as it outperformed Claude Code, which holds a 22.99% benchmark on the same task. While the broader PostTrainBench research recorded a high of 33% using the larger Gemma-3-4B model, the ability of ml-intern to extract 32% performance from the much smaller 1.7B parameter model demonstrates a significant advancement in automated training efficiency.

Advanced Technical Strategies

Beyond standard fine-tuning, ml-intern utilizes sophisticated technical strategies to optimize model performance. In testing, the agent demonstrated the ability to generate synthetic data when existing datasets were insufficient for reliable fine-tuning. For example, in a healthcare-domain test, it wrote scripts to generate synthetic training examples focused on edge cases, such as multilingual emergency response scenarios and medical hedging language, before upsampling this data for training.
The agent also implements advanced reinforcement learning techniques, such as Group Relative Policy Optimization (GRPO). By using GRPO, the agent performs reinforcement learning from human feedback with lower memory overhead than standard PPO. During a math-domain test, the agent successfully launched training on A100 GPUs, monitored reward curves, and performed ablations to isolate effective components before finalizing the model checkpoint.