Saved in:
Bibliographic Details
Main Author: Yu, Fangyuan
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2502.17910
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913705869967360
author Yu, Fangyuan
author_facet Yu, Fangyuan
contents Modern language models rely on static vocabularies, fixed before pretraining, in contrast to the adaptive vocabulary acquisition observed in human language learning. To bridge this gap, we introduce vocabulary curriculum learning, an approach that improves pretraining efficiency with log-linear scaling gains relative to vocabulary size. Our method alternates between entropy-guided vocabulary expansion and model optimization, enabling models to learn transferable representations across diverse tokenization granularities. This approach naturally gives rise to an optimal computation allocation pattern: longer tokens capture predictable content, while shorter tokens focus on more complex, harder-to-predict contexts. Experiments on small-scale GPT models demonstrate improved scaling efficiency, reinforcing the effectiveness of dynamic tokenization. We release our code to support further research and plan to extend our experiments to larger models and diverse domains.
format Preprint
id arxiv_https___arxiv_org_abs_2502_17910
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Scaling LLM Pre-training with Vocabulary Curriculum
Yu, Fangyuan
Computation and Language
Artificial Intelligence
Modern language models rely on static vocabularies, fixed before pretraining, in contrast to the adaptive vocabulary acquisition observed in human language learning. To bridge this gap, we introduce vocabulary curriculum learning, an approach that improves pretraining efficiency with log-linear scaling gains relative to vocabulary size. Our method alternates between entropy-guided vocabulary expansion and model optimization, enabling models to learn transferable representations across diverse tokenization granularities. This approach naturally gives rise to an optimal computation allocation pattern: longer tokens capture predictable content, while shorter tokens focus on more complex, harder-to-predict contexts. Experiments on small-scale GPT models demonstrate improved scaling efficiency, reinforcing the effectiveness of dynamic tokenization. We release our code to support further research and plan to extend our experiments to larger models and diverse domains.
title Scaling LLM Pre-training with Vocabulary Curriculum
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2502.17910