Conversation
4a69959 to
eb21dab
Compare
|
Trying to think if we have better options here... For For STORM, since that's solver-specific and doesn't need to be installed by default for the evals, we can probably drop that dep for astabench and maybe kick the issue to agent-baselines? I'm inclined to say that users of agent-baselines could be instructed to do the Publishing these to pypi is maybe not the end of the world but doesn't feel like an intended workflow and I worry that there might be unforeseen headaches as a result. Lmk if those suggested alternatives seem harder though, in which case maybe we should just go forward with this. |
|
👍 to kicking the STORM can over to Using a submodule is a decent alternative. I'll try that instead. |
https://github.com/allenai/astabench-issues/issues/283
PyPI does not allow git dependencies. Publish our own versions of
inspect_evalsandknowledge-stormso we can declare them as dependencies. Make a workflow file for each, which contains the specific git sha we will publish from. Need to manually override the name/version, which is done by making a static copy of the pyproject.toml/setup.py. Add docs for this procedure.Not pretty! The only alternative I can think of is to fork the repos, modify the pyproject.toml in the fork, and add the workflow files there. I prefer this approach (slightly) because it keeps all of the activity in this repo. However, if we ever have another project that depends on those same sha-specific dependencies, then publishing from the fork seems like the only option.