AI4Debunk knowledge graph and retrieval dataset for fact-checked claims

# AI4Debunk knowledge graph and retrieval dataset for fact-checked claims **License** CC BY-NC-SA 4.0 (see [LICENSE](LICENSE)) · **Access** Restricted Contains the knowledge graph, vector retrieval store, multimedia, and evaluation artifacts. ## Contents - `kg/` -- 2,826,776 RDF triples in 5 named graphs (gzipped N-Quads). - `vectors/` -- text and image indices as Parquet. - `multimedia/` -- 13,756 images (902 WP6 + 12,854 MOCHEG). - `metadata/` -- provenance tables and model registry with pinned HuggingFace revisions. - `evaluation/` -- 100 queries, 13,594 relevance labels, cached system outputs, expected results. - `SHA256SUMS` -- integrity checksum for every file. System (d) proposed in the deliverable reports P@10 = 0.249, R@10 = 0.279 (Horvitz-Thompson), F1@10 = 0.228. ## Attribution **Data sources.** CimpleKG (Burel et al., CKG@ISWC 2024; CC BY-NC-SA 4.0, snapshot 2026-05-18), AI4Debunk WP6 corpus from `AI4Debunk/climate-change-disinformation-detection-3rdrelease` and `AI4Debunk/war-in-ukraine-disinformation-detection_3rdrelease` on HuggingFace (CC BY 4.0), MOCHEG (Yao et al., SIGIR 2023; CC BY 4.0). **Models.** ReLiK for entity linking and closed information extraction (Orlando et al., ACL 2024; Apache 2.0), BGE-large-en-v1.5 and bge-reranker-v2-m3 (Xiao et al., SIGIR 2024; MIT), BM25 via Qdrant/bm25 in FastEmbed (Robertson & Zaragoza, 2009; Apache 2.0), Jina CLIP v2 (Jina AI, 2024; CC BY-NC 4.0), SSCD (Pizzi et al., CVPR 2022; CC BY-NC 4.0), EasyOCR (Apache 2.0), Llama-3.1-8B-Instruct (Dubey et al., 2024; Meta Llama 3.1 Community License). **Methods.** Reciprocal Rank Fusion for retrieval fusion (Cormack et al., SIGIR 2009), Horvitz-Thompson estimator for unbiased recall (Horvitz & Thompson, 1952), hybrid human + LLM pooling for relevance judgements (Otero & Parapar, SIGIR 2026), stratified sampling of the assessment pool (Webber et al., CIKM 2010).