A local-model tutorial for selecting reserved candidate-token logits, normalizing a complete choice distribution, and tuning that distribution directly with multiclass Brier loss.
Original by liushiliushiGetting startedGitHub repositorySource reviewed
Before you dive in
What you’ll find in the original
Map every allowed choice to a reserved single-token decision slot before normalization.
Train the complete decision distribution rather than only maximizing the winning label.
Use multiclass Brier loss when probability quality is part of the objective, then evaluate calibration separately.
Worth knowing
JevTuner applies ConfTuner's idea to closed-set decisions. It does not claim to reproduce TypeSafe's unpublished RLCD loss or architecture, and the README supplies no benchmark proving calibration.