Post has attachment
Szymon Sidor commented on a post on Blogger.
It's awesome that you presented a case for model being able to invent new captions. Is there a more extensive analysis available? What percentage of captions are copied VS invented? Do those percentages change if we condition on caption being mostly correct? I know it's hard to define a metric here, but potentially word overlap can lead to sensible results.
no plus ones
Wait while more posts are being loaded