Friday 21 March 2025
The notion of computational notebooks has revolutionized the way data scientists and researchers work. These interactive documents allow users to write code, perform calculations, and visualize results in a seamless, self-contained environment. However, despite their widespread adoption, notebooks have been plagued by issues of non-executability, where many notebooks are unable to run due to errors or compatibility problems.
A recent study has shed new light on this problem, revealing that the traditional notion of executability – requiring a notebook to run fully and without error – is overly rigid. The researchers found that many notebooks are not entirely broken, but rather suffer from partial executability issues that can be addressed with suitable execution environments.
The study analyzed a dataset of over 42,000 public notebooks, containing more than 34,000 non-executable notebooks. Using large language models (LLMs) and machine learning techniques, the researchers were able to identify and categorize notebooks into two main groups: potentially restorable and pathological non-executable notebooks.
Potentially restorable notebooks are those that can be fixed with relatively simple changes, such as updating dependencies or modifying code syntax. These notebooks may have been created using older versions of software or libraries, or may contain errors that can be easily corrected.
Pathological non-executable notebooks, on the other hand, are those that are fundamentally flawed and cannot be restored to a working state. These notebooks often contain complex dependencies or are tied to specific hardware configurations that make it difficult or impossible to reproduce their execution.
The researchers found that while 76% of public notebooks are currently considered non-executable, only about 7,000 of these notebooks are truly pathologically broken. This means that many more notebooks can be restored with the right tools and techniques.
To address this issue, the authors developed a novel approach using LLMs to generate synthetic input data for restoring non-executable notebooks. This technique was able to fully restore around 5% of previously non-executable notebooks, while partially improving the executability of an additional 12.8%.
The implications of these findings are significant. By recognizing that many notebooks are not entirely broken, researchers and developers can focus on creating tools and techniques to address the specific issues affecting each notebook. This could lead to a major increase in the reusability and reliability of computational notebooks.
Moreover, this study highlights the potential of LLMs in improving the overall quality and usability of notebooks.
Cite this article: “Reviving Non-Executable Notebooks: A Novel Approach Using Large Language Models”, The Science Archive, 2025.
Computational Notebooks, Non-Executability, Executability, Partial Executability, Execution Environments, Large Language Models, Machine Learning, Restorable Notebooks, Pathological Non-Executable Notebooks, Data Science Research
