# Hugging Face — Jupyter Agents: training LLMs to reason with notebooks

- Company: Hugging Face (huggingface.co)
- Announced: 2025-09-10T00:00:00+00:00
- Category: research-paper
- Subject: Platform
- Models affected: Qwen-3 Coder, Qwen3-4B-Thinking-2507, Claude 4 Sonnet, Qwen3-4B
- Source: https://huggingface.co/blog/jupyter-agent-2
- Record: https://forck.live/items/1626-jupyter-agents-training-llms-to-reason-with-notebooks

Hugging Face introduces Jupyter Agents, a project to train small LLMs to reason with notebooks for data science tasks. They fine-tune Qwen3-4B on Kaggle notebooks and evaluate on DABStep benchmark, achieving improved accuracy after scaffolding optimization.

## Evidence

Verbatim from https://huggingface.co/blog/jupyter-agent-2:

> accuracy jumped from 44.4% → 59.7% (easy split).

---

Record: https://forck.live/items/1626-jupyter-agents-training-llms-to-reason-with-notebooks
Catalogue: https://forck.live/llms.txt
Feed: https://forck.live/feed.md