FlavorReasonBench-Coffee: Benchmark Dataset

FlavorReasonBench-Coffee is a benchmark for evaluating large language model reasoning over professional coffee flavor hierarchies. 275 questions across 9 task types: T1.1–T1.5: hierarchical structure tasks (root classification, ancestor verification, sibling identification, path reconstruction, and common ancestor). T2.1–T2.3: similarity judgment tasks (ranking, pairwise comparison, odd-one-out). T3: open-ended professional coffee scenarios (LLM-as-judge scoring). Questions are grounded in a 1,175-node flavor graph derived from the SCA Coffee Taster's Flavor Wheel. The included tool graph (coffee_flavor_wheel.json) is the 111-node subset used as a structured tool in the original study. See README.txt for file formats, task descriptions, scoring rules, and evaluation notes.