You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains the reference solutions of the MettleBench benchmark. Access is gated to keep solutions out of web crawls and training corpora. Please describe briefly how you intend to use them (e.g. validating the benchmark, error analysis). Do not redistribute, publish, or use these files as training data.
Log in or Sign Up to review the conditions and access this dataset content.
MettleBench — reference runs (gated)
For each of the 82 MettleBench tasks this dataset provides:
reference/<task_id>/forward_only_solution.sh— a forward-only reference solution: a linear script, no rollback and no retry, that satisfies every acceptance criterion in a single pass;reference/<task_id>/order_violating_run.sh— the same assignment carried out in an order that ignores the dependencies among the criteria, which therefore stops at an earlier criterion;- any auxiliary files those scripts copy from.
Both take the workspace to operate on as their first argument and are meant to be run on a fresh copy of the shipped workspace:
cp -r tasks/<task_id>/workspace /tmp/ws
bash reference/<task_id>/forward_only_solution.sh /tmp/ws
python tasks/<task_id>/verifier/verifier.py --workspace /tmp/ws --output r.json
With this dataset downloaded, validate.py --reference in the GitHub
repository (Kelvin-Coffee/MettleBench) reproduces the three
admission checks of every task.
Please do not redistribute these files or use them as training data.
Authors
- Kelvin Coffee (@KelvinCoffee / GitHub)
- Yu Zhuang (@Zhuang1231 / GitHub)
- Downloads last month
- 19