Data and code for: Benchmarking nine large language models against human consensus for drug-related problem coding

Model outputs, human rating tallies and analysis code behind a benchmark of nine large language models against a pharmacist consensus standard for drug-related problem coding. The deposit contains 1,980 model sessions with verbatim responses, the per-cell vote tallies of two back-to-back pharmacist rating rounds on constructed vignettes, the reference code sets, and the scripts that regenerate every reported statistic and figure. The reference standard is a consensus product rather than an expert panel: 139 and 46 pharmacists rated 20 and 15 vignettes independently, and a cell is treated as consensus only when its reference code set reaches the k-th highest vote. The AI test set comprises 98 consensus cells and 7 contested cells. Data are released under CC BY 4.0; code under the MIT licence.