_A_akhe·Apr 18, 2024
Very nice, thanks for sharing! I noticed their token sequence length is fixed:
Constructing 7-gram LM ...
Created trie with 21,502,513 nodes.
People on here will be happy to say that I do a similar thing, however my sequence length is dynamic because I also use a 2nd data structure - I'll use pretentious academic speak: I use a simple bigram LM (2-gram) for single next-word likeliness and separately a trie that models all words and phrases (so, n-gram). Not sure how many total nodes because sentence lengths vary in training data, but there are about 200,000 entry points (keys) so probably about 2-10 million total nodes in the default setup.
"Constructing 7-gram LM": They likely started with bigrams (what I use) which only tells you the next word based on 1 word given, and thought to increase accuracy by modeling out more words in a s
Read the rest
equence, and eventually let the user (developer) pass in any amount they want to model (https://github.com/google-research/google-research/blob/5c87...). I thought of this too at first, but I actually got more accuracy (and speed) out of just keeping them as bigrams and making a totally separate structure that models out an n-gram of all phrases (e.g. could be a 24-token long sequence or 100+ tokens etc. I model it all) and if that phrase is found, then I just get the bigram assumption of the last token of the phrase. This works better when the training data is more diverse (for a very generic model), but theirs would probably outperform mine on accuracy when the training data has a lot of nearly identical sentences that only change wildly toward the end - I don't find this pattern in typical data though, maybe for certain coding and other tasks there are those patterns though. But because it's not dynamic and they make you provide that number, even a low number (any phrase longer than 2 words) - theirs will always have to do more lookup work than with simple bigrams and they're also limited by that fixed number as far as accuracy. I wonder how scalable that is - if I need to train on occasional ~100-word long sentences but also (and mostly) just ~3-word long sentences, I guess I set this to 100 and have a mostly "undefined" trie.
I also thought of the name "LMJS", theirs is "jslm" :) but I went with simply "next-token-prediction" because that's what it ultimately does as a library. I don't know what theirs is really designed for other than proving a concept. Most of their code files are actually comments and hypothetical scenarios.
I recently added a browser example showing simple autocomplete using my library: https://github.com/bennyschmidt/next-token-prediction/tree/m... (video)
And next I'm implementing 8-dimensional embeddings that are converted to normalized vectors between 0-1 to see if doing math on them does anything useful beyond similarity, right now they look like this:
[nextFrequency, prevalence, specificity, length, firstLetter, lastLetter, firstVowel, lastVowel]
For example, here's the embedding for "dog":
[0.6666666666666666, 0.6, 0.45714285714285713, 0.15, 0.12, 0.24, 0.56, 0.56]
where it used to just be 1D:
{ dog: 23 }
Up until now I've only effectively used nextFrequency to predict the next word, but with these other 7 metrics I'm able to store more info about each word and being vectors am able to store it efficiently, and do what all the smart people keep saying like dot product similarity index etc. The big boys are using like 1536D arrays etc. but with just 7 it still works, and it's weird that it works.
Thanks again for sharing!