Source-Aware Cross-Lingual Corpus of IP License Agreements from SEC EDGAR and Korean DART Filings

An open, cross-lingual corpus of intellectual-property license-agreement records reconstructed from public regulatory disclosures: U.S. SEC EDGAR 10-K filings (fiscal 2019–2025) and Republic of Korea DART business reports (fiscal 2023–2026). Contents. 38,114 raw extraction records; 19,054 quality-filtered records (SEC 13,951 + DART 5,103) covering 2,186 unique companies; 17,497 financial-term rows including 2,046 royalty-rate observations linked to quality-filtered agreements. Scope spans patents, software, know-how, and trademarks under a single schema. Provenance. Every record carries the identifier of the filing it was extracted from: 98.9% of extraction rows and 99.0% of quality-filtered rows resolve to a SEC accession number or a DART receipt number. Rows whose source could not be resolved uniquely are labelled in source_resolution_status (edgar_accession_ambiguous, edgar_accession_not_found) rather than silently dropped, so users can see exactly which records are not traceable. Files. license_contracts.csv, license_contracts_clean.csv, financial_terms.csv, companies.csv, filings.csv, together with dataset_summary.json, source_resolution_audit.json, schema_manifest.json, DATA_DICTIONARY.md, README.md, CITATION.cff, the CC BY 4.0 license text, and the extraction prompts and keyword lists. Use boundary. Validation converges at 76–79% genuine-license precision. The corpus supports exploratory, distributional, and first-pass benchmarking use with source verification. It is not suitable as sole evidence for legal due diligence, valuation opinions, transfer-pricing documentation, or record-level reliance of any kind. Supersedes v1.0.0 and v1.0.2, which carried empty source-identifier columns.