chunked/nltk_chunking/model_doc_clip/chunk_2.txt

It can be
instructed in natural language to predict the most relevant text snippet, given an image, without directly optimizing
for the task, similarly to the zero-shot capabilities of GPT-2 and 3.