AIFaultBench is a benchmark of 770 real-world AI software faults collected from 105 open-source repositories across 76 organizations, spanning traditional machine learning, deep learning, large language model infrastructure, reinforcement learning, agentic AI systems, and AI tooling. Each fault ships with the original GitHub issue report, a minimal reproduction script, a dependency specification, codebase reconstruction and environment setup scripts, reproduction logs, structured metadata, and a reproduction trajectory. 652 of the 770 faults (85%) are verified reproducible; the remainder document the reasons preventing reproduction.

If you use AIFaultBench, please cite it as below.