Benchmark Dataset for Environmental Values and Sustainability Alignment in Large Language Models

This repository contains a benchmark dataset for evaluating environmental values, sustainability-related attitudes, and behavioural recommendations in large language models (LLMs).

The dataset was developed to systematically assess environmental cognition, affect, and behavioural orientations expressed by LLMs across a broad set of sustainability-related prompts. It includes responses from 31 widely used proprietary and open-weight models evaluated under multiple prompting conditions.

The benchmark combines:

• questions derived from established environmental awareness surveys,
• sustainability-related behavioural measures,
• multilingual prompt formulations,
• comparative model evaluations,
• and derived sustainability-related indices.

The repository includes:

• a consolidated Excel workbook,
• machine-readable CSV exports for all sheets,
• YAML prompt definitions,
• documentation for index construction,
• and metadata intended to support FAIR and reproducible research practices.

The dataset enables:

• comparative benchmarking of LLM sustainability alignment,
• analysis of environmental attitudes embedded in model outputs,
• investigation of contextual sensitivity and persona-based steering effects,
• and comparison between LLM responses and human survey benchmarks from Germany.

The benchmark is intended as a reusable framework for future research on AI governance, sustainability-related value alignment, steerability, and normative robustness in generative AI systems.

Further methodological details and analysis will be provided in the corresponding research paper on arXiv: https://arxiv.org/abs/2606.02741

The repository with the code to generate this dataset can be found here: https://gitlab.opencode.de/uba-ki-lab/llm-questionnaire-benchmarking-framework