Perceptual conformity in multimodal language models follows Social Impact Theory: stimuli and per-trial measurements

Stimulus images and per-trial measurements for the article "Perceptual conformity in multimodal language models follows Social Impact Theory". Twelve open-weight vision-language models solve three perceptual tasks (line judgment, color recognition and dots estimation) and a fourth on handwritten digits, while the prompt reports the answers of a group of confederates; the probability of each answer is read from the logits of the two answer tokens. images.zip holds the 21,721 stimulus images of the three perceptual tasks, in 96 folders: for each of the 32 combinations of model and task, the pool of the main measurement, the difficulty set and the balanced pool. images_mnist.zip holds the 960 handwritten-digit stimuli, shared by the twelve models. results.zip holds the per-trial measurements of the eleven measurements reported in the article (2,049,252 trials), one file per model and task, with the prompt store of each measurement. The archives are unpacked into the root of the code repository, https://github.com/alebellina412/conformity_LLMs, whose scripts/download_zenodo.sh downloads and checks them; the manifests of the repository list every image with its MD5 checksum, and src/reduce.py turns results/ into the reduced tables from which every figure and table of the article is computed. Version 2 replaces the files of version 1: stimulus names without generation parameters, only the stimuli that meet the accuracy criterion of the analysis, and the per-trial measurements in place of intermediate outputs. Licences: CC BY 4.0, except images_mnist.zip, whose images derive from the MNIST dataset and are distributed under CC BY-SA 3.0, the licence of MNIST.