On moral role models and the Sustainable Development Goals: data and code (VATE v2 benchmark, raw generations, pilot archive)

Data and code for the manuscript On moral role models and the Sustainable Development Goals: a pre-registered test of role-model prompting in small language models (D. D. Hromada, Berlin University of the Arts; in preparation, not peer reviewed). Telling a language model to be "inspired by" admired figures is a cheap way to give it values. This pre-registered study tests whether it improves judgement or only changes direction: 8 open-weight models answered Sustainable Development Goal dilemmas whose option texts are identical across two versions while the stated facts favour the ecological option in one and the alternative in the other (25,600 responses, 8 system-prompt conditions). data_and_code.zip: the VATE v2 benchmark (68 evidence-conditioned dilemmas, 34 trade-off dilemmas, 30 adversarial probes), presentations and prompt conditions, pre-registration hash file, all raw per-generation records (JSONL with prompt, template and dataset hashes), judge key-validation and justification ratings, human annotation sheets (answer keys withheld for blind validation), and the generation, analysis and judge scripts. pilot_archive.zip: the unpublished pilot (VATE v1 items, scorers and prompt-search runs) from which the pilot-selected prompts used in hypothesis H4 were taken.