Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling

TL;DR

This paper introduces Over-Tokenized Transformers, a framework that decouples input and output vocabularies, showing that larger input vocabularies improve language model performance and scaling laws.

Contribution

It proposes a novel decoupled vocabulary approach that scales input vocabularies with multi-gram tokens, improving performance without extra costs.

Findings

Larger input vocabularies lead to lower training loss.

A log-linear relationship exists between input vocabulary size and performance.

Using large vocabularies achieves comparable results to bigger models.

Abstract

Tokenization is a fundamental component of large language models (LLMs), yet its influence on model scaling and performance is not fully explored. In this paper, we introduce Over-Tokenized Transformers, a novel framework that decouples input and output vocabularies to improve language modeling performance. Specifically, our approach scales up input vocabularies to leverage multi-gram tokens. Through extensive experiments, we uncover a log-linear relationship between input vocabulary size and training loss, demonstrating that larger input vocabularies consistently enhance model performance, regardless of model size. Using a large input vocabulary, we achieve performance comparable to double-sized baselines with no additional cost. Our findings highlight the importance of tokenization in scaling laws and provide practical insight for tokenizer design, paving the way for more efficient and powerful LLMs.