Programmatically verifiable coding tasks generated from vLLM git commits for the ControlArena vLLM setting.
This dataset contains 30 coding tasks automatically generated from git commits in the vLLM repository. Each task represents a real-world coding challenge derived from actual development work.
23 tasks (76%) are programmatically verifiable with pytest tests that can automatically validate task completion.… See the full description on the dataset page:
https://huggingface.co/datasets/RoganInglis/vllm-verifiable-control-arena.
