When pretraining at 8B scale, SuperBPE models consistently outperform the BPE baseline on 30 downstream tasks (+8% MMLU), while also being 27% more efficient at inference time.🧵
When pretraining at 8B scale, SuperBPE models consistently outperform the BPE baseline on 30 downstream tasks (+8% MMLU), while also being 27% more efficient at inference time.🧵
The result: sperm whale vowels do not just look like human vowels. They also behave like them.
We found several parallels. Like in Latin, whales have short and long vowels.
What is BPE?
• BPE is a subword tokenization algorithm
• It breaks text into frequently occurring character patterns
• Used in modern NLP models (GPT, BERT, etc.)
github.com/piyushdev04/...
What is BPE?
• BPE is a subword tokenization algorithm
• It breaks text into frequently occurring character patterns
• Used in modern NLP models (GPT, BERT, etc.)
github.com/piyushdev04/...
-40 bpe into hmc tho, stay tuned for a fatty pb
-40 bpe into hmc tho, stay tuned for a fatty pb
If you are looking a weekend project, here you go: sebastianraschka.com/blog/2025/bp...
If you are looking a weekend project, here you go: sebastianraschka.com/blog/2025/bp...
computer, use tiktoken to BPE tokenize & embed this Pomeranian
computer, use tiktoken to BPE tokenize & embed this Pomeranian
This was my final PhD degree requirement :)
The goal was to unify the major tokenization algorithms under a finite-state automaton framework. For example, by encoding a BPE tokenizer as a transducer.
This was my final PhD degree requirement :)
The goal was to unify the major tokenization algorithms under a finite-state automaton framework. For example, by encoding a BPE tokenizer as a transducer.
www.usatoday.com/story/opinio...