Jupyter Agents: training LLMs to reason with notebooks

Jupyter Agents enhance LLMs with code execution capabilities inside Jupyter notebooks, enabling them to solve data science tasks without leaving the workflow. Through a pipeline of dataset curation, QA generation, and model fine-tuning, researchers improved Qwen3-4B's performance on the DABStep benchmark from 38.7% to 75% accuracy on easy tasks, demonstrating that even small models can become effective data analysis agents with proper training and scaffolding.