title: OO-LD
emoji: 🔗
colorFrom: gray
colorTo: gray
sdk: static
pinned: false
Object-Oriented Linked Data
Schemas that are JSON Schema and JSON-LD at once, so one document validates and carries its own semantics. oo-ld.org
What is here
Schema corpora. The generated modules a benchmark reads: quantity kinds from QUDT, and classes from schema.org. Pinned by digest rather than by version, because two generations of the same corpus can differ in ways that change an answer.
Adapters. LoRAs for schema-constrained extraction, each carded with what it was measured at and in which condition it is worth having. The short answer is: only where the prompt withholds the class catalogue. Where the catalogue is present, every one of them scores below the model it was trained from.
Results. The graded cells behind the published figures.
The question they answer
Where does structure have to be enforced for a model to produce data that validates: in the prompt, at decode time, at commit time, or nowhere. The code, the method and the findings are at github.com/OO-LD/oold-llm-bench.