The first release of the DMTN-337: Performance of Machine-Learned Reliability Scoring for Image Differencing is ready and available here: https://dmtn-337.lsst.io/
DMTN-337 reports on the status of the ML reliability (Real/Bogus) model for Rubin. It describes the model architecture and why it was chosen. The document defines the model versioning scheme for present and future reference and recaps all existing model versions, starting with the model used for DP1 data, through the model for DP2 data, to the one currently being used to deliver alerts. It describes the data used for each version, including how the labels obtained from the initial release of the Zooniverse Citizen Science Project - ‘Rubin Difference Detectives’ were analyzed and used. The document also defines the reliability cutoff, currently applied to all alerts, by presenting how each model version behaves with fake injections and a random subset of recent LSSTCam detections, and presents purity and completeness curves for different datasets.
We welcome your questions and feedback in this forum topic.