CHinese Homophonic Associative Reasoning with Multimodality Benchmark.
CHARM is a multimodal benchmark inspired by a Chinese homophonic puzzle game. Each problem includes two images, a hint word, an answer category, and an answer length. Models can interact multiple times with the environment by submitting guesses. After each incorrect guess, the environment returns feedback on character positions and pinyin positions, which helps the model… See the full description on the dataset page:
https://huggingface.co/datasets/YuanNang/CHARM-Bench.
