BPE from Scratch

A ground-up Python implementation of Byte-Level Byte Pair Encoding (BBPE) — the tokenization algorithm used by GPT-2, GPT-3, GPT-4, and other modern LLMs.

What is Byte-Level BPE?

BPE was originally a data compression algorithm that replaces the most frequent pair of bytes in a sequence with a single unused symbol. Applied to NLP, it builds a subword vocabulary by iteratively merging the most frequent adjacent token pairs in a corpus.

The byte-level variant (introduced by OpenAI for GPT-2) operates directly on raw UTF-8 bytes rather than characters or words:

• Base vocabulary of 256 — one token per possible byte value (0–255), no unknown tokens ever
• Language-agnostic — any Unicode text (code, math, emoji, CJK, ...) is representable without a special <UNK> token
• Lossless — encoding and decoding are exact roundtrips
• Merges learned greedily — at each step, the most frequent adjacent pair is merged and assigned a new token ID (256, 257, ...)

This is the same fundamental approach used by tiktoken (OpenAI) and Hugging Face tokenizers for GPT-style models.

Algorithm Phases

Installation

pip install bpe-from-scratch

Usage

Train from scratch

from bpe_from_scratch import ByteLevelBPE

bpe = ByteLevelBPE()
bpe.train(text, vocab_size=1024)  # 1024 total tokens, 768 merge rules
bpe.save("my_model.json")

Or train directly from a folder of .txt files:

from bpe_from_scratch.train import train_from_folder

bpe = train_from_folder(
    folder_path="data/corpus_A/",
    model_path="my_model.json",
    vocab_size=1024,
)

Utilize all CPU cores

Pass num_workers=os.cpu_count() to parallelize pre-tokenization across all cores (useful for large corpora):

import os
from bpe_from_scratch import ByteLevelBPE

bpe = ByteLevelBPE()
bpe.train(text, vocab_size=50_257, num_workers=os.cpu_count())

Or via the folder helper:

import os
from bpe_from_scratch.train import train_from_folder

train_from_folder(
    folder_path="data/corpus/",
    model_path="my_model.json",
    vocab_size=50_257,
    num_workers=os.cpu_count(),
)

Note: On Windows, guard the call site with if __name__ == "__main__": due to spawn-based multiprocessing. On macOS/Linux, fork is used by default and no guard is needed.

Encode and decode

tokens = bpe.encode("Hello, world!")  # list[int]
text   = bpe.decode(tokens)           # "Hello, world!"

Continue training on new data

Load an existing model and extend the vocabulary without discarding what was already learned:

bpe = ByteLevelBPE()
bpe.load("my_model.json")
bpe.continue_train(new_text, new_vocab_size=1280)  # extend to 1280 total tokens
bpe.save("my_model.json")

All previously learned token IDs remain stable — documents encoded with the old model are still valid after the update.

Or use the folder helper:

from bpe_from_scratch.train import continue_train_from_folder

bpe = continue_train_from_folder(
    folder_path="data/corpus_B/",
    model_path="my_model.json",
    new_vocab_size=1280,
)

Note: continue_train uses a frozen-base approach — existing merges are replayed on the new text before new rules are learned. This keeps token IDs stable but is not equivalent to a full retrain on the combined corpus. See TRAINING_GUIDE.md for details and tradeoffs.

Project Structure

src/bpe.py        # Core implementation
tests/test_bpe.py # Unit tests
tests/manual/     # Interactive notebooks for experimentation

Running Tests

PYTHONPATH=src python3 -m unittest discover -s tests -v

Acknowledgements

Inspired by Andrej Karpathy's minbpe.

References

• minbpe — Minimal BPE implementation by Andrej Karpathy
• GPT-2 Paper — Language Models are Unsupervised Multitask Learners (Radford et al., 2019)
• Byte-Pair Encoding tokenization — Hugging Face NLP Course
• Neural Machine Translation of Rare Words with Subword Units — original BPE-for-NLP paper (Sennrich et al., 2016)
• Building a Fast BPE Tokenizer from Scratch — Jun Yu Tan — Stages 1–5 complexity analysis and benchmarks
• From Hours to Seconds: Optimising BPE Tokeniser Training — Logan Thomson — Stages 1–7 including GC reduction and adaptive parallelism
• Bypassing the GIL for Parallel Processing — Real Python — multiprocessing vs threading guidance