This is a glitch token [1]! As the article hypothesizes, they seem to occur when a word or token is very common in the original, unfiltered dataset that was used to make the tokenizer, but then removed from there before GPT-XX was trained. This results in the LLM knowing nothing about the semantics of a token, and the results can be anywhere from buggy to disturbing. A common example is usernames that participated on…
Aren’t there only 2^16 tokens? Seems easy to test for all of them, but I might just not understand the tokenizer.
From what I've found through Google (with no real understanding of llm) 2^16 is the max tokens per minute for fine tuning OpenAI's models via their platform. I don't believe this is the same as the training token count.
Then there's the context token limit, which is 16k for 3.5 turbo, but I don't think that's relevant here.
Though somebody please tell me why I'm wrong, I'm still trying to wrap my head around the training side.