Decoupling the Benefits of Subword Tokenization for Language Model Training via Byte-level Simulation

TL;DR

This paper investigates how subword tokenization improves language model training by isolating its effects in a byte-level setting, revealing key factors like throughput and boundary priors.

Contribution

It introduces a controlled byte-level pretraining pipeline to decouple subword effects, providing new insights into their roles in model performance and efficiency.

Findings

Subword boundaries act as explicit priors or biases enhancing performance.

Increased training throughput significantly benefits subword models.

Byte-level simulation clarifies the specific contributions of subword tokenization.

Abstract

Subword tokenization is an essential part of modern large language models (LLMs), yet its specific contributions to training efficiency and model performance remain poorly understood. In this work, we decouple the effects of subword tokenization by isolating them within a controlled byte-level pretraining pipeline. We formulate and test hypotheses across various dimensions, including sample throughput, vocabulary scaling, and the linguistic prior of subword boundaries. By simulating these effects in a byte-level setting, we refine our understanding of why subword models outperform raw byte models and offer insights to improve the pretraining of future byte-level and subword models. Specifically, our experiments highlight the critical role of increased training throughput and the integration of subword boundaries as either explicit priors or inductive biases.