Compute Optimal Tokenization

TL;DR

This paper investigates how tokenization granularity impacts language model scaling laws, revealing that optimal compression rates differ from BPE and vary with compute, influencing tokenization choices for efficiency.

Contribution

It systematically studies the effect of token compression rate on scaling laws using 988 models, showing parameter count scales with data size in bytes, not tokens.

Findings

Model parameter counts scale proportionally to data size in bytes.

Optimal compression rate decreases with increased compute.

Findings apply across tokenization methods and languages.

Abstract

Scaling laws enable the optimal selection of data amount and language model size, yet the impact of the data unit, the token, on this relationship remains underexplored. In this work, we systematically investigate how the information granularity of tokens, controlled by the compression rate (i.e., average bytes of text per token), affects scaling trends. We train 988 latent tokenized models (BLT) ranging from 50M to 7B parameters that enable setting the desired compression rate. This flexibility allows us to study the role of compression rate well beyond 4.57 bytes per token obtained with a popular BPE tokenizer. Our experiments reveal that in compute-optimal configurations, model parameter counts scale proportionally to data size measured in bytes, not in tokens as commonly perceived (Kaplan et al., 2020; Hoffmann et al., 2022). Furthermore, we discover that the optimal compression rate differs from the one obtained with BPE and decreases with compute. These findings generalize to both latent and subword tokenization, as well as to languages other than English, guiding language model developers on tokenization scheme selection for maximal compute efficiency.