Maintainer workflows

Contributing and verification

• Read the contribution guide before opening a pull request.

• Use Testing for the fast local loop and full verification commands.

• Use Development for internal API and compatibility-shim conventions.

• Use Orchestrator architecture when changing worker processes, hardware planning, or training data residency.

Training and model releases

Maintained operational documentation lives next to the scripts it describes:

• Training pipeline

• Release, synchronization, and deployment

• Generated model and data bundles

• Release training recipe records the data and component settings used for the published 2.3.0 weights.

• Neural training semantics explains framework equations and compatibility choices.

• Reproduce the 2.3.0 model training reruns the recorded 2.3.0 recipe from a clean checkout. A new fit need not reproduce the original training trajectory.

The public release entry point is:

mhcflurry train pan-allele-release --help

It coordinates training, evaluation, plots, remote artifact synchronization, and optional deployment. Deployment is never enabled by default.

Training internals

These pages specify current behavior of the training and evaluation pipeline. Read Release training recipe first for the component definitions the others reference.

• Affinity checkpoint comparison covers --save-all-checkpoints and comparing terminal against minimum-validation-loss weights.

• Processing data preparation describes how mhcflurry train processing-data builds affinity-matched negatives with resumable artifacts.

• Processing negative matching specifies the seeded without-replacement negative-matching contract.

• Presentation percentile calibration documents the shared CompactPercentRankTransform budget for presentation calibration.

• Evaluate saved model runs evaluates a completed training run with mhcflurry eval saved-candidate, without retraining.

Controlled experiments

These pages define controlled comparisons: the question, the fixed controls, and the commands that run them.

• Cleavage-boundary processing models asks whether boundary-spanning sequence context adds signal after controlling for binding affinity.

• Processing kernel-width experiment sweeps kernel widths across the legacy flank CNN and the boundary-branch architecture.

• Processing hyperparameter campaign compares processing-specific training recipes rather than assuming affinity’s settings transfer.

• Replaying historical processing data replays historical processing data against the actual public affinity weights.