[TACL'23] VSR: A probing benchmark for spatial undersranding of vision-language models. - View it on GitHub