Project
Glossary
Use consistent terms across chat, embeddings, manifests, training, and release operations.
- Canonical record: provider/model-neutral semantic data with provenance.
- Codec: external-format conversion with an explicit loss report.
- Template: late-bound tokenizer-specific chat rendering.
- Prompt convention: late-bound embedding query/document instruction or prefix.
- Response distillation: transfer of generated assistant responses.
- Vector distillation: transfer of embedding geometry.
- Score/ranking distillation: transfer of relevance values or ordering.
- Frozen dataset: immutable records, splits, contamination evidence, rights, and hashes.
- Full resume: restoration of training state, not merely weights.
- Warm start: weights-only initialization.
- Recipe lock: exact model, revision, license evidence, conventions, capabilities, and gates.
- Supported: every declared gate passed for the stated scope.
- Experimental: usable for bounded evaluation without production stability.
- Unavailable: capability fails closed because required evidence is absent or failed.