A focused visual interpreter that takes images and produces natural language descriptions. It handles the bootstrapping problem of vision-language learning by filtering noisy web image-text pairs before training, which gives its captions more coherent grounding than models trained on raw scraped data. Works well for straightforward scene description but can struggle with fine-grained spatial relationships or rare object categories.