This 8B model was trained from scratch with a SuperBPE tokenizer. SuperBPE extends the BPE algorithm to include both traditional subword tokens (contained within word boundaries), as well as new superword tokens (containing parts of multiple words)! Due to encoding the same amount of text in fewer tokens, this model is 30% more efficient at inference-time on average compared to a model trained with BPE.
The model was trained with the Olmo2 7B architecture and pretraining data. It has a context length of 2,884 tokens (to match the effective context size in bytes of a BPE model with a context length of 4,096 tokens), and is trained on 332B tokens. The tokenizer has a vocabulary size of 200k and transitions from learning subword to learning superword tokens at vocabulary size of 160k.
Example Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("UW/OLMo2-8B-SuperBPE-t160k")
model = AutoModelForCausalLM.from_pretrained("UW/OLMo2-8B-SuperBPE-t160k")
tokenizer.convert_ids_to_tokens(tokenizer.encode("By the way, I am a fan of the Milky Way."))
# ['ByĠtheĠway', ',ĠIĠamĠa', 'ĠfanĠofĠthe', 'ĠMilkyĠWay', '.']
Citation
@misc{liu-etal-2025-superbpe,
title={SuperBPE: Space Travel for Language Models},
author={Alisa Liu and Jonathan Hayase and Valentin Hofmann and Sewoong Oh and Noah A. Smith and Yejin Choi},
year={2025},
eprint={2503.13423},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2503.13423},
}
