chunked/nltk_chunking/model_doc_imagegpt/chunk_4.txt

Despite training on low-resolution ImageNet without labels,
we find that a GPT-2 scale model learns strong image representations as measured by linear probing, fine-tuning, and
low-data classification.