Outputs and prediction provenance

Keep the model, effective configuration, preprocessing metadata and split identifiers together. Predictions alone are not enough to reproduce a run or determine which rows were held out.

Main artifacts

Artifact	Purpose
xirt_model.pt	Final PyTorch model checkpoint for compatible 2.0 inference
configuration.json	Effective network and learning configuration
split_NN.json	Recorded training, validation, and prediction IDs for fold NN
error_features.csv	Predictions and per-task prediction-error features when table writing is enabled
error_features_interactions.csv	Derived interactions between error features when table writing is enabled
encoder.json	Exported alphabet, padded sequence length, and fraction-level mapping
Callback logs/checkpoints	Optional fold-level training history and checkpoints according to callback settings
mlflow.db, mlflow-artifacts/	Output-local tracking store and artifacts when tracking is enabled without overriding paths

Some artifacts depend on mode, task, preprocessing route, and the --write setting. ESM feature caches are separate from recurrent label-encoder metadata. The exact model architecture and feature/preprocessing contract must be compatible when reusing a checkpoint.

Checkpoints retain _sequence_length, _target_scales, the vocabulary and fraction mappings, input-content provenance, and the selected model's per-fold _split. Prediction on a smaller subset reuses the saved sequence width and target scale rather than shrinking padding or choosing a new scale from that subset.

The source-content fingerprint hashes a documented serialization of the input table; it is not necessarily the raw CSV file checksum. Keep the original file checksum as an additional provenance record when needed.

callbacks.check_point controls writing optional fold checkpoints. Best-epoch selection/restoration also uses an in-memory state and does not require those files to be written. The final xirt_model.pt is not a TensorFlow HDF5 model; legacy HDF5 files require isolated 1.x inference or retraining, not renaming.

Interpret the cv column

Value	Interpretation
Positive fold number	Predictions assigned to that fold's held-out prediction IDs; inspect the recorded split and run mode
0	Final-model predictions populated for training/validation rows; not out-of-fold estimates
-1	Predictions for the separate non-training candidate set

For an out-of-fold comparison, use only genuinely held-out fold predictions and verify the sample IDs against split_NN.json. Positive fold numbers alone do not prove a leakage-free scientific split. A single train/test run is not equivalent to full crossvalidation, and rows used for validation are not an independent final test set.

Do not combine cv=0 predictions with out-of-fold predictions and report the combined error as crossvalidated accuracy. Likewise, candidate-set predictions do not establish accuracy without appropriate independent observed labels.

Reproducibility and privacy

Archive input checksums/identifiers, task labels and units, transformation settings, source revision, package versions, seeds, and hardware with the outputs. Keep paired-peptide reorderings traceable through identification IDs.

MLflow fingerprints training features and targets and logs configuration/metrics/model artifacts; it does not upload feature arrays or labels. A fingerprint does not replace the data/split provenance needed to reproduce a result. Model weights and configuration remain potentially sensitive and should not be added to public CI artifact paths indiscriminately.

Consult the benchmark protocol for matched comparisons and the migration guide before reusing legacy models.