You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
1. Adds a new script `scripts/dl_katago_networks.py` which downloads all the KataGo networks to `/nas/ucb/k8/go-attack/katago-networks`. The script will detect if a network has already been downloaded and skips downloading in this case.
2. Add a new eval to `scripts/generate_paper_evaluations.py` which runs the an adversary against the latest KataGo networks. The adversary as well as the earliest KataGo network to evaluate against in configurable in `scripts/paper_evaluations.yml`. The new eval can be configured to either run on a CHAI machine or Hofvarpnir.
3. Cleans up the `scripts/generate_paper_evaluations.py` file to remove the need to launch a devbox to generate configs. This is slow and also unnecessary (one can instead generate configs on either a CHAI machine or Hofvarpnir by setting up a python environment using something like micromamba).
4. Remove the `codecov` dependency to make the CI work. See #113.
See b15a92b for an example of the actual eval configs that were generated.
See https://www.notion.so/chaiberkeley/Evals-against-latest-KataGo-checkpoints-1d11bd9115c347c49e21939eda833ba2?pvs=4 for experiments that use this new eval script.
See https://about.codecov.io/blog/message-regarding-the-pypi-package/
The text was updated successfully, but these errors were encountered: