A dataset of 100 visual RPG combat conflict samples designed to evaluate Vision-Language Models (VLMs) under cross-modal conflicts (discrepancy between battle screenshots and caption text). Derived from the rcannizzaro/rpg_fantasy_battle_counterfactual_v2 dataset.

Dataset Statistics

This dataset consists of a single train split containing 100 perfectly isolated conflict samples derived from the… See the full description on the dataset page: https://huggingface.co/datasets/multilingual-vlm-conflict/rpg-conflict.