mli0118/ontology-llm

This repository implements a benchmark for detecting taxonomic hallucinations in a language model, comparing an Algebraic Ontology Projection (AOP) detector with confidence and consistency baselines. The intended workflow is to derive labeled probes from an ontology, obtain model answers and hidden-state representations, evaluate detectors, and report results. The sampled modules establish ontology-to-probe generation, model I/O, AOP training/evaluation, and baseline calls. benchmark.py and the result artifact are present in the tree but were not sampled, so their orchestration and data links are not asserted.