tokenmonster VS sentencepiece

The actual analysis: https://github.com/alasdairforsythe/tokenmonster/blob/main/b...

> Summary of Findings:

> - Comparable (50256-strict-nocapcode) TokenMonster vocabularies perform better than both GPT-2 Tokenizer and tiktoken p50k_base on all metrics.

> - Optimal vocabulary size is 32,000.

> - Simpler vocabularies converge faster but do not necessarily produce better results when converged.

> - Higher compression (more chr/tok) does not negatively affect model quality alone.

> - Vocabularies with multiple words per token have a 5% negative impact on SMLQA (Ground Truth) benchmark, but a 13% better chr/tok compression.

> - Capcode takes longer to learn, but once the model has converged, does not appear to affect SMLQA (Ground Truth) or SQuAD (Data Extraction) benchmarks significantly in either direction.

> - Validation loss and F1 score are both meaningless metrics when comparing different tokenizers.

> - Flaws and complications in the tokenizer affect the model's ability to learn facts more than they affect its linguistic capability.

I need to benchmark the performance of my tokenizer against standard tokenizers. It would be best for reproducibility if I benchmark against an existing model on a standard benchmark, swapping out the existing tokenizer for my tokenizer.

Having just released my ungreedy subword tokenizer (TokenMonster), I keep being ask to provide benchmarks on how it performs when actually used to train a model, vs other tokenizers.

This is all I've been doing 16 hours per day, 7 days per week for the past couple of months. If you like it please ☆ star the GitHub so people will find it. If you have any questions feel free to ask on here or on the GitHub Discussions tab. Thank you.

TokenMonster is an ungreedy tokenizer and vocabulary builder, outperforming tiktoken by 35%. In fact, TokenMonster's smallest 24000 vocabulary consistently uses less tokens than tiktoken's largest 100256 vocabulary to tokenize the same text. Save the tokens! See benchmark.

From the GitHub:

Code at Github.

you need to train the model on 1 trillion tokens (https://platform.openai.com/tokenizer https://github.com/google/sentencepiece) anyways for it to get reasoning capacities, which it feels very unlikely that your data is that much.

I'm highly skeptical that you have enough data to pretrain if you don't have enough data to fine tune.

fine tuning + vector search + prompting of as much stuff as you can, on a LLM like palm2 or gpt4 is what I would do. otherwise you can use falcon 40B ofc.

maybe I should charge for this ahah

a) Comparison with SentencePiece tokenizer with comparable settings (It can also ignore word-boundaries and create phrase tokens)

LLaMA uses the sentencepiece tokenizer: https://github.com/google/sentencepiece

Besides, are you familiar with SentencePiece? What you are doing looks very similar (generate a large vocab, prune worst token until vocab size is reached), only the token selection criterion is different. It's also purely data driven in the sense that there are no assumption specific to language (and it can optionally segment across whitespace, as you are doing).

Hi, I'm studying the implementation of encode and decode functions for Google's SentencePiece tokenizer.

I think many LMs nowadays use unicode tokenizers, that are not tied to specific languages. E.g. sentencepiece is the most popular one: https://github.com/google/sentencepiece

If you are interested in implementing LLaMA yourself or learning, I noticed that the reference code by Facebook is one of the cleaner, easier to read ML code I've seen in a while. https://github.com/facebookresearch/llama/blob/main/llama/mo... It's about 200 lines long. You probably do need a bit of knowledge to understand what you are reading but I was pleasantly surprised.

For example in comparison, StableDiffusion torch code in diffusers and transformers Python libraries has lots of conditionals, experiments etc. that are not being used that can make it hard to follow what is going on.

Last weekend I got the "main loop" of the transformer working in pure CPU Rust code, following the reference code. My crappy code is just very very slow as I focused on getting it to run, not making it fast. The tokenizer uses some Google thing https://github.com/google/sentencepiece but luckily for inference it seems that you just need to be able to parse the tokenizer model file and not understand how it was created; I was able to strip out the protobuf files from that repository and add it to Rust and read the tokens.

I am optimistic that someone makes a high quality CPU or some CPU+GPU+SSD combination thingmaling that will make it somewhat practical to run even the large LLM models without needing an A100 or two.