Per-task workspace tarballs for MettleBench, a benchmark of 82
long-horizon engineering tasks for evaluating LLM agents.
Workspaces are distributed as tarballs because several contain git
repositories, symbolic links, and permission bits that need to be
preserved exactly.
This dataset holds only the data. The prompts, evaluators, download and
validation tooling, and documentation live in the GitHub repository:
Kelvin-Coffee/MettleBench. Use its… See the full description on the dataset page:
https://huggingface.co/datasets/KelvinCoffee/MettleBench.
