JupOtter: Cell-Level Bug Detection in Jupyter Notebooks

A new system called JupOtter targets the growing problem of bugs in Jupyter Notebooks, the interactive coding environment widely used in data science, by introducing a cell-level detection method that its creators say outperforms existing tools on most tested benchmarks. The system, detailed in a paper posted to the arXiv preprint server on June 22, 2026, is built around three core components: a tokenization strategy that preserves a notebook’s cell structure, a technique for predicting bugs at the individual cell level, and a new annotated dataset named OtterDataset [1][2]. OtterDataset contains more than 21,000 notebooks labeled for fine-grained, cell-level bug detection [1][2]. Jupyter Notebooks, which blend code, text, and visualizations in a single document, have moved beyond simple prototyping. The authors note that notebooks are “increasingly used to develop more complex programs, leading to a rapid rise in buggy notebooks on platforms like GitHub” [2]. This shift has created demand for debugging tools that understand the non-linear execution and cell-based architecture unique to notebooks, rather than treating them like traditional scripts. In evaluations across three datasets, JupOtter achieved cell-level bug detection F1 scores that surpassed both static analyzers and large language models on two of the three [1][2]. The paper does not provide performance figures for the third dataset. Large language models, which are machine learning models with many parameters trained on vast amounts of text for tasks such as language generation, have been applied to a range of code-related problems, but the results suggest they are not yet optimized for the specific structure of computational notebooks [8]. The research appears on arXiv, an open-access repository that hosts electronic preprints in fields including computer science and statistics [6]. As of late 2024, the repository was receiving about 24,000 new articles per month, and it surpassed two million total articles at the end of 2021 [6]. The paper is accompanied by arXivLabs integrations, a framework launched in 2020 that allows community collaborators to build experimental tools on top of the repository’s article pages [4]. These integrations, which appear as tabs on the abstract page, include citation explorers and code-finding services, and are governed by guidelines that require partners to adhere to arXiv’s values of openness and user data privacy [4][5]. The JupOtter paper’s abstract page displays several such tools, including the Bibliographic Explorer for navigating citation trees and the CORE Recommender for surfacing related open-access research [5]. arXivLabs is currently pausing new proposals while the development team focuses on migrating the platform’s systems to the cloud, though existing projects are unaffected [3].