Assessing Scientific Knowledge in Patents: A Large Language Model Approach (dataset and prompts)

This repository contains the data release accompanying the paper “Assessing Scientific Knowledge in Patents: A Large Language Model Approach.” The paper was published in Quantitative Science Studies (2026) (https://doi.org/10.1162/QSS.a.493). It provides patent identifiers and derived large language model (LLM) outputs used in the analyses, as well as the prompt texts. Files - `LLM_outputs.csv` Patent-level LLM scores and model-generated explanations.- `prompts.txt` files. Full prompt texts used in the study. Unit of observation Each row in `llm_patent_science_scores.csv` corresponds to one **patent × model × prompt variant × patent parts × iteration** evaluation. Columns (LLM_outputs.csv) - `patent_number`: USPTO patent identifier.- `model_name`: Model/checkpoint name used for scoring.- `prompt_variant`: Prompt condition/variant label.- `patent_parts`: Patent sections provided to the model (e.g., title/abstract/claims).- `iteration`: Repeated run identifier (integer).- `score`: Science-knowledge score (1 = technology-driven, 10 = science-driven).- `model_explanation`: Model-generated textual justification for the assigned score. Notes - This repository does **not** include patent full texts; it contains patent identifiers and derived LLM outputs. License Creative Commons Attribution 4.0 International (CC BY 4.0).