Retrained the LightGBM model on a fresh corpus extracted from eFa archives. Initial test had AUC 0.998 but after a config change it dropped massively. Suspect feature extraction mismatch.
1 comment
Transcom
13 Apr 2026, 09:40
extract-features-v2.py wasn't mirroring the scanner's featurize() function exactly — drift in one-hot encoding for content-type. Synced both, back to 0.998 AUC.
1 comment
Sign in with your TDesk account to comment.