DepDoctor: A Dataset of Python Code Broken by Dependency Upgrades, Paired with Fixes

DepDoctor is an open dataset of 6,204 examples of Python code that breaks when a library is upgraded, each paired with the corrected code and a short note describing the API change that caused the break. It mixes real cases mined from permissively licensed public GitHub commits with synthetic cases generated from a database of known breaking changes, and it includes 935 no-op examples where the correct action is to leave the code unchanged, to teach a model restraint. The dataset is intended for training and evaluating models that repair Python code after a dependency upgrade. A Qwen2.5-Coder-7B model fine-tuned on this data reaches 62 percent fully correct fixes on a held-out 100-example test set. Full evaluation, data fields, collection method, cleaning, and licensing are described in the included documentation. Each real code row keeps the permissive license of its source repository, recorded in the license field of that row. A live, loadable version is hosted on the Hugging Face Hub.