Extracting Structure-Activity Relationships from Patents with Large Language Models

This dataset contains all patent documents used for training and validating the prompts described in the associated research article Extracting Structure-Activity Relationships from Patents with Large Language Models. Including files: PatSAR-Dev-50.zip: 50 patents used for prompt development only. PatSAR-Test-214.zip: 214 PARP-related patents used for core evaluation. PatSAR-GenTest-200.zip: 200 patents covering ten non-PARP small-molecule targets used for generalization evaluation. DocSAR-Patent-96.zip: 96 patent PDFs extracted from the DocSAR-200 benchmark for external comparison, where permitted by the source dataset. Consistency subset: 20 previously evaluated patents, with 10 from PatSAR-Test-214 and 10 from PatSAR-GenTest-200.