benchflow/posttrain-generic-dogfood-20260922

PostTrain generic-runner dogfood collection (2026-09-22)

A clearly labeled test collection for PostTrain Arena's generic hosted execution path (contract bash-script-v2): the model writes each task's solve script and the package's own oracle and verifier score it. It repackages three existing public starting-kit examples as one team entry. It is not a newly authored benchmark, a competition-scale corpus, or evidence of held-out improvement.

Source and attribution

• Original project: https://github.com/benchflow-ai/posttrainarena (AGPL-3.0)
• Immutable source commit: 1b9091a64997a1ad829e81172536d30109d22711
• envs/dogfood-hello-text, envs/skillsbench-weighted-gdp-calc, and envs/skillsbench-3d-scan-calc are unchanged copies of starting-kit/examples/<name> at that commit. Original authors are retained in each task.md, which also declares its license, origin and a category from PostTrain Arena's list, so the arena can credit each task.
• This packaging adds only this README and the required submission.yaml.

Local checks

python3 scripts/check_task.py /path/to/this-entry/envs
scripts/run_local.sh /path/to/this-entry/envs/dogfood-hello-text
scripts/run_local.sh /path/to/this-entry/envs/dogfood-hello-text --skip-oracle

Structural validation and local replay do not measure post-training performance.