| name | model-release-inference |
| description | Use when packaging a TeraiNet model release, building or reviewing inference code, exporting class mappings, validating model artifacts, or preparing deployment to Hugging Face or AddaxAI. |
Model Release and Inference
Release Bundle
Package a retained model with all information required to load and use it without training data:
- model file and format;
- ordered class-list artifact;
- effective configuration and preprocessing contract;
- source data and split provenance;
- seed, framework dependency versions, and checkpoint-selection rationale;
- validation, test, and out-of-distribution metrics reported separately.
Use terainet.models.create_class_list_yaml_file() to create the class-list artifact rather
than recreating label-index logic in a notebook or deployment script.
Inference Verification
- Load the model and class-list artifact in a clean process with no training dataset access.
- Validate input image mode, resize, normalization, channel order, and batch shape against the
training preprocessing contract.
- Confirm each output index maps to the saved ordered class list.
- Run deterministic predictions on representative held-out examples and record results.
- Treat missing or incompatible metadata as a release blocker, not a warning.
Deployment Boundaries
- Do not expose credentials, access tokens, or local filesystem assumptions in release code.
- Do not claim generalization from validation data or merge OOD results into test metrics.
- Version model artifacts and configuration together; never overwrite a published artifact in
place without an explicit new release identifier.
Completion Checklist