• Create a well-documented model zoo. [priority:: low] [completion:: 2025-07-30]

A model zoo has to be created to “remember” models that are trained on some data, and how they perform. Ideal metrics are RMSE, which provide a numerical sense, however need to take the scaling into account.

Turns out we can actually use Gitlab as an MLFlow repository, so

  • Investigate use of GitLab model registry as model zoo [priority:: medium] [completion:: 2024-10-23]

Seem to get 404 errors when connecting to `curl -H “PRIVATE-TOKEN: ${MLFLOW_GITLAB_API_TOKEN}” “https://gitlab.cern.ch/api/v4/projects/195476/mlflow

Querying mattermost channel to see if anyone knows why I cannot access the mlflow endpoint

  • Follow up on 404 errors on GitLab model registry [priority:: medium] [completion:: 2024-10-24]

  • Evaluate models prior to upload to sps-models-hysteresis-compensation to associate model experiments with some metric [priority:: medium] [completion:: 2025-07-30]

  • Decide on a naming convention for model uploads It turns out that I have no control over the model versioning, and the model version is not tied to a model version candidate / artifact freely. If we don’t specify the model name, a random name is assigned. We should name the models [model name]_v[version], and then a script can check for the existence by querying model candidates and looking for the latest version with a regular expression. [completion:: 2024-11-21]

Model naming convention

  1. Create a model for each architecture and purpose. E.g. TFT.SPS.MBI or TFT.SPS.SIM.