justinchuby/onnx-genai-inference-metadata-catalogue

Private ONNX inference metadata catalogue

catalogue.jsonl contains one JSON object per private real-weight package or canonical metadata profile. Rows pin source and package revisions, source license, runtime evidence, output summaries, and any substitution rationale. Model repositories contain canonical hashless inference_metadata.yaml; artifact integrity hashes live only in distribution evidence and manifests.

<!-- inference-metadata-annotations:start -->

Collection annotation audit

All 28 model repositories in the ONNX GenAI Inference Metadata Examples collection publish inference_metadata.annotated.yaml beside canonical inference_metadata.yaml, with a model-card link. Comments explain the actual workflow, profiles, tensor/cache/state/batching contracts, capability obligations, authority boundaries, and fail-closed omissions for each package.

The 12 complete-generation examples use schema v1.2 token authority: package-default numeric IDs live only under package.tokenizer.special_tokens, while request EOS tensors are replacement overrides consumed by executable termination. Logits-only examples intentionally do not claim package-managed sampling or termination.

• —`annotation_inventory.json` pins every current model revision, identifies the canonical token-authority migrations, and records verification status.
• —`validate_annotations.py` inventories the collection at execution time, verifies parsed semantic equality, checks comments, README/provenance links and hashes, classifies generation scope, and validates both YAML files with the current onnx-genai validator.

The dataset itself is the collection catalogue, not a model package, so it has no canonical inference_metadata.yaml to annotate. <!-- inference-metadata-annotations:end -->