Automatically Detecting Under-Trained Tokens in Large Language Models
1–10 of 27 posts
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#2Good Computerphile video on glitch tokens a year ago:
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#3Amazing name for the paper
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#4Isn't the solution to just train the tokeniser on the same corpus as the LLM? I'm not sure why reusing tokenisers is so common. Anybody know?
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#5Isn't the solution to just train the tokeniser on the same corpus as the LLM? I'm not sure why reusing tokenisers is so common. Anybody know?
From the abstract I get the feeling these techniques are useful when you don’t have access to the corpus, as e.g. in the case where you download some open source weights but the corpus is secret. Otherwise I don’t understand why you wouldn’t just compute a histogram over the tokens in (a statistical sample of) the corpus.
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#6Isn't the solution to just train the tokeniser on the same corpus as the LLM? I'm not sure why reusing tokenisers is so common. Anybody know?
Sure, but if your corpus is very large, that's not feasible.
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#7Isn't the solution to just train the tokeniser on the same corpus as the LLM? I'm not sure why reusing tokenisers is so common. Anybody know?
From the abstract I get the feeling these techniques are useful when you don’t have access to the corpus, as e.g. in the case where you download some open source weights but the corpus is secret. Otherwise I don’t understand why you wouldn’t just compute a histogram over the tokens in (a statistical sample of) the corpus.
> open source weights but the corpus is secret
This is oxymoronic; the corpus is the "source". Yet this usage of "open source" is widespread. Maybe we should start calling such models by their rightful name, "freeware".
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#8Isn't the solution to just train the tokeniser on the same corpus as the LLM? I'm not sure why reusing tokenisers is so common. Anybody know?
Sure, but if your corpus is very large, that's not feasible.
Tokenizer training doesn't scale as well as model training, so general practice is to train on a subset of the full corpus.
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#9Isn't the solution to just train the tokeniser on the same corpus as the LLM? I'm not sure why reusing tokenisers is so common. Anybody know?
From the abstract I get the feeling these techniques are useful when you don’t have access to the corpus, as e.g. in the case where you download some open source weights but the corpus is secret. Otherwise I don’t understand why you wouldn’t just compute a histogram over the tokens in (a statistical sample of) the corpus.
The paper mentions some reasons why these quick fix ideas are not as simple as it sounds. For example many rare tokens are “intermediate” merges inside the BPE algorithm, shorter prefixes of longer words. The long word is common, but its earlier, intermediate merge is not, by itself.
Re: Automatically Detecting Under-Trained Tokens in Large Language Models
#10Good Computerphile video on glitch tokens a year ago: https://www.youtube.com/watch?v=WO2X3oZEJOA
This video somehow looks even more interesting than the pre-print of the article