MetaHotpotQA

MetaHotpotQA is a schema-aware multi-hop Knowledge Graph Question Answering (KGQA) benchmark that bridges natural-language multi-hop questions with structured Wikidata reasoning. Starting from the 12,056 questions in this dataset are sourced from HotpotQA (Yang et al., 2018) and grounded into Wikidata through an ontology-guided pipeline built on top of the 10 domain ontologies from Text2KGBench (Mihindukulasooriya et al., 2023), covering Movie, Music, Book, Sport, Military, Computer, Space, Politics, Nature, and Culture. Each record is annotated with:- Wikidata QIDs for seed entities and the answer entity- A difficulty label (entity_selection, traversal, property_comparison)- A reasoning type (bridge, comparison)- Matched ontological domains- Recovered supporting paths in the per-question Wikidata subgraph (where applicable) The dataset is split into train (8,439), dev (1,809), and test (1,808) using stratified sampling over difficulty × reasoning_type with a fixed random seed (42). Answer coverage is guaranteed by construction: only questions whose answer can be structurally located in the KG are retained. This deposit contains:- published_splits.zip: the canonical train/dev/test JSONL files (~7 MB)- subgraphs_dataset.tar.gz: per-question ontology-filtered Wikidata subgraphs organised by split, one JSON file per question (~528 MB compressed). These subgraphs are the source from which the oracle KG context is constructed for evaluation. Baseline experiments on the test split (1,808 questions) with four 4-bit quantized LLMs (Qwen3-14B, Gemma3-12B, Llama3.1-8B, Phi3-14B) show a mean +48.5 pp Exact Match gain of oracle KG over closed-book, validating that the KG structure provides genuine reasoning signal beyond parametric knowledge. Construction code, evaluation scripts, and baseline predictions are available in the accompanying GitHub repository.