FoodSafeSum

FoodSafeSum is a machine-actionable dataset for NLP in food safety. It contains human-written and LLM-generated summaries and titles of 2,091 food-safety documents, plus manually curated topics, document types, and automatically extracted hazard annotations. Documents were gathered by SGS Digicomply from news, regulatory/legal sources, guidance portals, and scientific outlets (years 2002–2023; ~58% originally in English; the rest translated and curated). The dataset enables research on classification, retrieval, RAG-style QA, and event clustering in food-safety monitoring and policy. (See Section 3 and Table 7 for schema and fields; Figures 1–3 for source/type statistics. In the paper) What’s included? Manual summary and manual title (by domain experts) LLM summary and LLM title (generated with meta.llama3-70b-instruct via Bedrock) Document type (News, Regulation, Guidance, Scientific) and topic labels (12 high-level categories, e.g., Policies & Laws; Contaminants, residues & contact materials) Hazard annotations auto-extracted from a controlled vocabulary derived from prior work Source name and original title For each source item: Note: The full original documents are not included in the public release (used internally for analysis only). Format & schema Primary release as CSV/JSON with columns (see Table 7): manual_summary, manual_title, llama70b_summary, llama70b_title, source_name, doc_type, topics, plus hazards (list) and any auxiliary metadata used for experiments. Multilingual inputs were translated (Google Translate/DeepL) and curated; see paper for details.