SWE-Misbehavior is a benchmark for measuring boundary violations made by software coding agents, meaning actions that go beyond what the user asked for. Examples include rewriting code around the target function, changing signatures, touching unrelated files, deleting the wrong files, running unrequested commands, or skipping files the agent needed to read. The benchmark has 7,427 task instances across 3 task families and 9 subsets. All of them are built on real-world Python repositories from DevEval (10 domains: Communications, Database, Internet, Multimedia, Scientific-Engineering, Security, Software-Development, System, Text-Processing, Utilities). Each instance comes with 50 paraphrased…
Independent publisher
Nguyen Duc Anh
duc-anh-2002
Optimization, Machine Learning, Generative AI
Models in Library0
Datasets in Library1
Models on Hugging Face—
Followers1
Compare the published evidence
This profile connects the 1 model and dataset record shown below to their upstream cards and sources. The library counts describe SAVRN's coverage, rather than this publisher's entire output.
Published tasks: Text generation.
Compare model specifications, source dates, and hardware estimates on the individual records. Publisher identity verification is separate from a model's quality, deployment readiness, or suitability for your work.