HuggingFace Develops Jupyter Agent

HuggingFace has built Jupyter Agent, which executes code directly within a Jupyter Notebook environment to enable LLMs to perform data analysis and data science tasks. This helps models carry out multi-step reasoning using code and markdown cells.

The project's main goal is to strengthen Small Models, which perform worse compared to large models. To this end, the team is focused on generating high-quality training data and building a pipeline that fine-tunes models to improve performance on relevant benchmarks.

The key strategies for improving performance are as follows:

• Realistic data science task evaluation via the DABStep benchmark
• Scaffolding optimization: simplifying the existing complex framework to effectively control model behavior
• Fine-tuning experiments targeting small models such as Qwen3-4B

In practice, restructuring by simplifying the scaffolding resulted in a significant improvement in accuracy on Easy tasks for the Qwen3-4B model, from 44.4% to 59.7%.