German General Social Survey Personas – GGSS Personas Add-On data

The German General Personas is a dataset based on the survey data from the German General Social Survey (GGSS/ALLBUS) wave of 2023. The data converted the tabular survey data into a natural language format that is compatible with the use in Large Language Model (LLM) prompts. We designed a process for the systematic selection of survey variables featured in different versions of the GGP. Since the idea of the GGP is to be a general-purpose persona prompt collection, this persona attribute selection process aims at identifying the subsets of sociodemographic and additional attributes that are most important for explaining variation in human responses across a range of different topics. The GGP data comprises a total of 20 versions, differing in format (key-value pairs and full-text versions) and level of included information (ten levels, ranging from only core socio-demographic information to all available survey variables). Each version includes all 5,246 participants surveyed in the GGSS/ALLBUS 2023 as persona descriptions, featuring varying numbers of the variables available in the GGSS/ALLBUScompact 2023 (ZA8831). The first version of the persona description format is a JSON-like key-value structure, where the questions or statements in the ALLBUScompact are the keys and their response options in text format are the corresponding values. To ensure that the response options (the variable labels are usually stored as integers in the tabular survey data) are interpretable for LLMs, we converted them back into the original text labels found in the survey codebook and interview question documentation. In addition to the fully structured key-value persona descriptions, we also provide more descriptive natural-language versions of the same personas. Each of these persona descriptions is created by prompting a proprietary, state-of-the-art LLM (Gemini-2.5-flash-lite) to turn the key-value persona description into a short textual description without adding, altering or omitting any of the provided information.