Character Prefix Conditioning(cursor.com)
cursor.com
Character Prefix Conditioning
https://www.cursor.com/blog/cpc
7 comments
> Can you construct an efficient algorithm for sampling from q(tk∣t1,…,tk−1), that minimizes calls to the original language model?
I feel like I'm missing some issue here... Can't you query stopping at the last full token boundary, then reject any results which don't match the character prefix and continue from there with the completion? Kind of like when you mask the invalid actions when reinforcement training on games? Or is that losing too much info?
I feel like I'm missing some issue here... Can't you query stopping at the last full token boundary, then reject any results which don't match the character prefix and continue from there with the completion? Kind of like when you mask the invalid actions when reinforcement training on games? Or is that losing too much info?
I asked o1 to figure this out and this is essentially what it came up with as well.
https://chat.sshh.io/share/HIzUotMYVxFhRde94ZYJJ
https://chat.sshh.io/share/HIzUotMYVxFhRde94ZYJJ
Not sure if this is free labor or a means to source candidates.
So here is ChatGPT's token list: https://gist.github.com/s-macke/ae83f6afb89794350f8d9a1ad8a0...
Is there some reason it isn't alphabetical? (More specifically, lexically sorted by codepoint) If you had a model with sorted tokens, you'd be able to solve this by constraining output to tokens with the desired prefix, probably with some mechanism similar to how this works: https://github.com/ggerganov/llama.cpp/blob/master/grammars/...
Is there some reason it isn't alphabetical? (More specifically, lexically sorted by codepoint) If you had a model with sorted tokens, you'd be able to solve this by constraining output to tokens with the desired prefix, probably with some mechanism similar to how this works: https://github.com/ggerganov/llama.cpp/blob/master/grammars/...
They are sorted, but in a way such that, roughly speaking, rank is proportional to inverse frequency (like Zipf’s law, but also permitting merging of subwords to be ranked). This is actually extremely important because it makes the otherwise very-high-cardinality categorical feature of target predicted argmax vocab dictionary key index slightly smoother and slightly more predictable for the model
516 instances of "\r\n"!
This is harder than it looks.
First "token-healing" doesn't work. Consider the case "app" where the most likely options are "ap|praisal" or "apple|sauce". You can't just sample all tokens that start with app, or you'd miss appraisal.
Second, it's easy to come up with a naive algorithm that samples from the true distribution. It's very difficult to make this algorithm efficient.
First "token-healing" doesn't work. Consider the case "app" where the most likely options are "ap|praisal" or "apple|sauce". You can't just sample all tokens that start with app, or you'd miss appraisal.
Second, it's easy to come up with a naive algorithm that samples from the true distribution. It's very difficult to make this algorithm efficient.
Ideally you'd have a language model that can predict a good continuation after any byte. If an existing model can't do that because it's too reliant on a specific tokenization, you might nonetheless be able to fine-tune it until it can gracefully handle the unexpected tokenizations that result from splitting at a random byte.
Such a model will always be less performant than one on tokens, as you're effectively switching to one byte per token. Solving this problem in code is much cheaper.
I don't mean switching to one byte per token, but switching to training on the token distribution that results from cutting off the input at arbitrary bytes. The bytes per token should be basically unchanged, as only the end gets a bit shorter.
Yeah I've tried that approach. The model ends up needing to learn every combination of tokens. For example, the word "apple" now has six bytes positions it can be split on and the model suddenly needs to learn that all six will yield the same output attention state.
It ends up being O(max token length) more complex and so you end up needing a proportionally larger model to accommodate it.
It ends up being O(max token length) more complex and so you end up needing a proportionally larger model to accommodate it.
Seems like we should just use gradual annealing of tokens to more fine grained single character tokens over the course of training then
I believe that's similar to the idea behind https://github.com/facebookresearch/blt
Do they have the solution (I'd assume Supermaven which they recently acquired does) and are they just giving out an interesting challenge to readers?
Or... what is this?
A related topic is "token healing", although some implementations (unfortunately including the one in HuggingFace Transformers) make some big assumptions that aren't always true (like treating spaces as special).