Code Generation
- HumanEval · subset · 2 measured runs
- HumanEval (full 164, serial greedy pass@1) · full · 5 measured runs
Task → benchmark → protocol → evidence
Nine stable capability areas. Domains, modalities, methods and safety properties are filters, not competing top-level categories.
OCR and document AI → · Measured open-weight models → · Taxonomy JSON
72 tasks imported from the original area map; 1 beta extension; 31 executable evaluation profiles. Imported tasks do not imply implemented adapters.