TREX: Tokenizer Regression for Optimal Data Mixture

TL;DR

TREX introduces a regression-based framework that efficiently predicts optimal data mixtures for multilingual tokenizer training, improving compression efficiency and scalability over heuristic methods.

Contribution

TREX presents a novel regression approach to predict optimal data mixtures for tokenizers, reducing reliance on heuristics and costly searches in multilingual LLM training.

Findings

TReX outperforms heuristic mixtures by up to 12% in compression efficiency.

The framework demonstrates strong scalability and robustness.

TReX enables scalable mixture search before large-scale tokenizer training.

Abstract

Building effective tokenizers for multilingual Large Language Models (LLMs) requires careful control over language-specific data mixtures. While a tokenizer's compression performance critically affects the efficiency of LLM training and inference, existing approaches rely on heuristics or costly large-scale searches to determine optimal language ratios. We introduce Tokenizer Regression for Optimal Data MiXture (TREX), a regression-based framework that efficiently predicts the optimal data mixture for tokenizer training. TREX trains small-scale proxy tokenizers on random mixtures, gathers their compression statistics, and learns to predict compression performance from data mixtures. This learned model enables scalable mixture search before large-scale tokenizer training, mitigating the accuracy-cost trade-off in multilingual tokenizer design. Tokenizers trained with TReX's predicted mixtures outperform mixtures based on LLaMA3 and uniform distributions by up to 12% in both inand out-of-distribution compression efficiency, demonstrating strong scalability, robustness, and practical effectiveness.