Post

There is no posttraining data commons

Open Athena argues that open model developers lack a shared supply of reliable agent-training tasks: unlike web documents, tasks need a reproducible environment, an objective, tools, and a way to grade outcomes.

The essay uses SWE-bench and other coding datasets to illustrate how missing dependencies, underspecified goals, broken tests and leaked solutions shrink the usable pool. It proposes deliberately building a public task commons rather than assuming traces alone are sufficient.

HN readers debated the usefulness of shared agent traces and raised consent, worker displacement and governance concerns; the author argued that a participatory commons could give workers more control than privately acquired datasets.