minbpe

Minimal, clean code for the (byte-level) Byte Pair Encoding (BPE) algorithm commonly used in LLM tokenization. The BPE algorithm is "byte-level" because it runs on UTF-8 encoded strings.

This algorithm was popularized for LLMs by the GPT-2 paper and the associated GPT-2 code release from OpenAI. Sennrich et al. 2015 is cited as the original reference for the use of BPE in NLP applications. Today, all modern LLMs (e.g. GPT, Llama, Mistral) use this algorithm to train their tokenizers.

There are two Tokenizers in this repository, both of which can perform the 3 primary functions of a Tokenizer: 1) train the tokenizer vocabulary and merges on a given text, 2) encode from text to tokens, 3) decode from tokens to text. The files of the repo are as follows:

1. minbpe/base.py: Implements the Tokenizer class, which is the base class. It contains the train, encode, and decode stubs, save/load functionality, and there are also a few common utility functions. This class is not meant to be used directly, but rather to be inherited from.
2. minbpe/basic.py: Implements the BasicTokenizer, the simplest implementation of the BPE algorithm that runs directly on text.
3. minbpe/regex.py: Implements the RegexTokenizer that further splits the input text by a regex pattern, which is a preprocessing stage that splits up the input text by categories (think: letters, numbers, punctuation) before tokenization. This ensures that no merges will happen across category boundaries. This was introduced in the GPT-2 paper and continues to be in use as of GPT-4. This class also handles special tokens, if any.
4. minbpe/gpt4.py: Implements the GPT4Tokenizer. This class is a light wrapper around the RegexTokenizer (2, above) that exactly reproduces the tokenization of GPT-4 in the tiktoken library. The wrapping handles some details around recovering the exact merges in the tokenizer, and the handling of some unfortunate (and likely historical?) 1-byte token permutations.
5. minbpe/llama.py: Implements the LlamaTokenizer (almost?). This class is a light wrapper around the RegexTokenizer (2, above) that implements following functionality found in LlamaTokenizer.
   1. BPE on unicode instead of UTF-8.
   2. GPT-4 regex pattern for splitting text, with splitting on individual digits instead of {1,3}.
   3. utf-8 fallback for bytes that are not in the tokenizer's unicode vocabulary.
   4. character coverage to handle rare tokens during training.
   5. ability to instantiate the tokenizer with a pre-trained model, using .from_pretrained() method.
   6. Notes/TODOs:
      1. The vocabulary starts with special tokens (4 in this case: unk, bos, eos, pad), followed by utf-8 fallback byte tokens (256), then unicode tokens (variable), and finally the merge tokens (variable).
      2. To use the tokenizer, first either train it on a corpus using the train method or load a pretrained one using the from_pretrained method, which can take a file saved using save method.
      3. Use encode_ordinary instead of encode to encode text. (#TODO)
      4. Added '<|pad|>' as a special token.
      5. Find the exact Llama regex pattern for splitting text. (#TODO)
      6. Haven't ported the code to train.py yet. (#TODO)
      7. Match the output with the output of LlamaTokenizer. (#TODO)
6. test/test_llama.py: Implements the tests for the LlamaTokenizer class, duplicating the tests for the GPT4Tokenizer class; however there were some differences to I made a new test file.

Finally, the script train.py trains the two major tokenizers on the input text tests/taylorswift.txt (this is the Wikipedia entry for her kek) and saves the vocab to disk for visualization. This script runs in about 25 seconds on my (M1) MacBook.

All of the files above are very short and thoroughly commented, and also contain a usage example on the bottom of the file.

quick start

As the simplest example, we can reproduce the Wikipedia article on BPE as follows:

from minbpe import BasicTokenizer
tokenizer = BasicTokenizer()
text = "aaabdaaabac"
tokenizer.train(text, 256 + 3) # 256 are the byte tokens, then do 3 merges
print(tokenizer.encode(text))
# [258, 100, 258, 97, 99]
print(tokenizer.decode([258, 100, 258, 97, 99]))
# aaabdaaabac
tokenizer.save("toy")
# writes two files: toy.model (for loading) and toy.vocab (for viewing)

According to Wikipedia, running bpe on the input string: "aaabdaaabac" for 3 merges results in the string: "XdXac" where X=ZY, Y=ab, and Z=aa. The tricky thing to note is that minbpe always allocates the 256 individual bytes as tokens, and then merges bytes as needed from there. So for us a=97, b=98, c=99, d=100 (their ASCII values). Then when (a,a) is merged to Z, Z will become 256. Likewise Y will become 257 and X 258. So we start with the 256 bytes, and do 3 merges to get to the result above, with the expected output of [258, 100, 258, 97, 99].

inference: GPT-4 comparison

We can verify that the RegexTokenizer has feature parity with the GPT-4 tokenizer from tiktoken as follows:

text = "hello123!!!? (안녕하세요!) 😉"

# tiktoken
import tiktoken
enc = tiktoken.get_encoding("cl100k_base")
print(enc.encode(text))
# [15339, 4513, 12340, 30, 320, 31495, 230, 75265, 243, 92245, 16715, 57037]

# ours
from minbpe import GPT4Tokenizer
tokenizer = GPT4Tokenizer()
print(tokenizer.encode(text))
# [15339, 4513, 12340, 30, 320, 31495, 230, 75265, 243, 92245, 16715, 57037]

(you'll have to pip install tiktoken to run). Under the hood, the GPT4Tokenizer is just a light wrapper around RegexTokenizer, passing in the merges and the special tokens of GPT-4. We can also ensure the special tokens are handled correctly:

text = "<|endoftext|>hello world"

# tiktoken
import tiktoken
enc = tiktoken.get_encoding("cl100k_base")
print(enc.encode(text, allowed_special="all"))
# [100257, 15339, 1917]

# ours
from minbpe import GPT4Tokenizer
tokenizer = GPT4Tokenizer()
print(tokenizer.encode(text, allowed_special="all"))
# [100257, 15339, 1917]

Note that just like tiktoken, we have to explicitly declare our intent to use and parse special tokens in the call to encode. Otherwise this can become a major footgun, unintentionally tokenizing attacker-controlled data (e.g. user prompts) with special tokens. The allowed_special parameter can be set to "all", "none", or a list of special tokens to allow.

training

Unlike tiktoken, this code allows you to train your own tokenizer. In principle and to my knowledge, if you train the RegexTokenizer on a large dataset with a vocabulary size of 100K, you would reproduce the GPT-4 tokenizer.

There are two paths you can follow. First, you can decide that you don't want the complexity of splitting and preprocessing text with regex patterns, and you also don't care for special tokens. In that case, reach for the BasicTokenizer. You can train it, and then encode and decode for example as follows:

from minbpe import BasicTokenizer
tokenizer = BasicTokenizer()
tokenizer.train(very_long_training_string, vocab_size=4096)
tokenizer.encode("hello world") # string -> tokens
tokenizer.decode([1000, 2000, 3000]) # tokens -> string
tokenizer.save("mymodel") # writes mymodel.model and mymodel.vocab
tokenizer.load("mymodel.model") # loads the model back, the vocab is just for vis

If you instead want to follow along with OpenAI did for their text tokenizer, it's a good idea to adopt their approach of using regex pattern to split the text by categories. The GPT-4 pattern is a default with the RegexTokenizer, so you'd simple do something like:

from minbpe import RegexTokenizer
tokenizer = RegexTokenizer()
tokenizer.train(very_long_training_string, vocab_size=32768)
tokenizer.encode("hello world") # string -> tokens
tokenizer.decode([1000, 2000, 3000]) # tokens -> string
tokenizer.save("tok32k") # writes tok32k.model and tok32k.vocab
tokenizer.load("tok32k.model") # loads the model back from disk

Where, of course, you'd want to change around the vocabulary size depending on the size of your dataset.

Special tokens. Finally, you might wish to add special tokens to your tokenizer. Register these using the register_special_tokens function. For example if you train with vocab_size of 32768, then the first 256 tokens are raw byte tokens, the next 32768-256 are merge tokens, and after those you can add the special tokens. The last "real" merge token will have id of 32767 (vocab_size - 1), so your first special token should come right after that, with an id of exactly 32768. So:

from minbpe import RegexTokenizer
tokenizer = RegexTokenizer()
tokenizer.train(very_long_training_string, vocab_size=32768)
tokenizer.register_special_tokens({"<|endoftext|>": 32768})
tokenizer.encode("<|endoftext|>hello world", allowed_special="all")

You can of course add more tokens after that as well, as you like. Finally, I'd like to stress that I tried hard to keep the code itself clean, readable and hackable. You should not have feel scared to read the code and understand how it works. The tests are also a nice place to look for more usage examples. That reminds me:

tests

We use the pytest library for tests. All of them are located in the tests/ directory. First pip install pytest if you haven't already, then:

$ pytest -v .

to run the tests. (-v is verbose, slightly prettier).

exercise

For those trying to study BPE, here is the advised progression exercise for how you can build your own minbpe step by step. See exercise.md.

lecture

I built the code in this repository in this YouTube video. You can also find this lecture in text form in lecture.md.

todos

• write a more optimized Python version that could run over large files and big vocabs
• write an even more optimized C or Rust version (think through)
• rename GPT4Tokenizer to GPTTokenizer and support GPT-2/GPT-3/GPT-3.5 as well?
• write a LlamaTokenizer similar to GPT4Tokenizer (i.e. attempt sentencepiece equivalent)

License

MIT