Data for "The psychophysics of affect in large language models"

Per-trial data for Marks & Vollberg, “The psychophysics of affect in large language models”. LLM-participants rated how gains and losses would feel (−50 to +50) and chose between gambles and sure outcomes, in Study 1 (abstract shape lottery, £ amounts) and Study 2 (stock trading, % price moves), under single-evaluation (NO_MEMORY), memory and joint conditions. The zip contains a data/ folder to place in the root of the code repository: the main batch (13 Study 1 and 6 Study 2 model × condition datasets; raw, quality-cleaned and per-trial annotated files, plus per-LLM-participant quality reports); the earlier Study 1 dataset in full, including the base model llama-3.1-405b (data/Study1_archive/); and the uncorrected originals of the files changed by the re-extraction corrections described in the Methods (data/_pre_reparse_backup/). Each trial row includes the model’s raw response alongside the extracted value. README.md describes the layout; SHA256SUMS.txt and MANIFEST.csv give checksums, row and byte counts.