More consistently follow instructions related to not OCRing text

[fpgaminer]

Instruction following for instructions that tell it not to read text (on watermarks, for example) is weak and/or requires very specific instructions.

[CognitiveDiffusion]

I just came here to comment this.

Beta1 is the best image captioner I ever used (been around since the original wd-tagger).

But... it sees watermarks everywhere. Even on a lot of images that don't feature any text.

[maedtb]

+1, it sees watermarks that do not exist, and it does not reliably follow directions to avoid mentioning watermarks. When I attempted to discourage these words via bad_word_ids, it very aggressively replaces letters/spaces with incorrect spellings and differently formed phrasings.

I've opted to just let it mention watermarks in its output and removed those sentences from the output in a post-process.