Jupyter Agent 2: Hugging Face improves agents that run notebooks

Hugging Face presents Jupyter Agent 2, a bet on agents that work directly with notebooks and datasets, execute code, and answer questions based on data. Can you imagine asking an agent to open a notebook, run an analysis, and give you the answer with reproducible code? That's exactly what they're trying to make easier. (huggingface.co)

Qué es Jupyter Agent 2

Jupyter Agent 2 is the next step in Hugging Face's effort to build agents that understand a notebook's context and can execute Python code (pandas, numpy, visualizations) to answer questions about datasets. It's not just text generation: the agent can produce step‑by‑step execution traces and calculated solutions designed for data analysis tasks. (huggingface.co)

The team also published the code and related demos in a Hugging Face Space that manages interactions with notebooks, making it easier to experiment without setting up the entire environment from scratch. (huggingface.co)

El dataset detrás: Jupyter Agent Dataset

To train these agents they launched the Jupyter Agent Dataset. According to Hugging Face, the dataset was built from real notebooks and Kaggle datasets, using a pipeline that filters content, generates question‑answer pairs, and produces executable traces that help the agent learn to reason with code. (huggingface.co)

Some key numbers: the team reports using terabytes of Kaggle data as a source, and the final dataset contains tens of thousands of synthetic notebooks (around 51k examples and roughly 2 billion tokens in the public subset). These examples combine real questions, verified answers, and automated code executions. (huggingface.co)

Resultados y entrenamientos

Hugging Face trained model variants on that dataset (for example, versions from the Qwen‑3 family) and reports measurable improvements on benchmarks focused on code generation and data analysis, with significant gains in metrics like DABstep on easier scenarios. That suggests exposing models to executable notebook traces improves their practical ability to solve data queries. (huggingface.co)

¿Qué puedes hacer con esto hoy?

• Try the agent demo in the Hugging Face Space to see how it answers questions about notebooks and datasets. (huggingface.co)
• Load the Jupyter Agent Dataset with datasets.load_dataset("data-agents/jupyter-agent-dataset") to train or evaluate your own agent. from datasets import load_dataset is exactly the starting point. (huggingface.co)

Imagine this applied to your workflow: an analyst uploads a CSV, the agent inspects columns, runs transformations, and returns charts and answers. Does it save you time? Yes. Does it fully replace human oversight? Not entirely, so it's wise to review the traces.

Precauciones y límites prácticos

Not everything is perfect: executable traces may not reproduce in every environment; LLM‑generated answers can contain errors and biases; and using material derived from third‑party notebooks means paying attention to licenses and terms of service like Kaggle's. Hugging Face recommends running code in safe sandboxes and validating results before making critical decisions. (huggingface.co)

Important: artifacts generated by LLMs can be wrong. Always validate operations that involve sensitive data or important decisions. (huggingface.co)

Cómo empezar rápido (pasos prácticos)

1. Visit the Jupyter Agent 2 Space demo to see examples in action. (huggingface.co)
2. If you want to train, download or load data-agents/jupyter-agent-dataset with the datasets library. ds = load_dataset("data-agents/jupyter-agent-dataset", split="non-thinking") is an example shown in the docs. (huggingface.co)
3. Try pretrained models or train a small model on your dataset to evaluate improvements on concrete data‑analysis tasks.

Reflexión final

Jupyter Agent 2 and its dataset are a practical step toward bringing AI into the notebook workflow: fewer abstract ideas and more ability to run real analyses. Does that mean the tools replace the analyst? No, but they do amplify what you can accomplish in less time. If you work with data, it's worth trying the demo and seeing whether these executable traces help your projects. (huggingface.co)