Achieving Tokenizer Flexibility in Language Models through Heuristic Adaptation and Supertoken Learning

TL;DR

This paper introduces Tokenadapt, a flexible tokenizer transplantation method with novel pre-tokenization learning for Supertokens, significantly improving language model efficiency and performance by reducing retraining and preserving semantics.

Contribution

The paper presents a model-agnostic tokenizer transplantation framework and a new pre-tokenization learning approach for Supertokens, enhancing compression and reducing fragmentation in language models.

Findings

Tokenadapt outperforms baseline methods like Transtokenizer and ReTok in initializing new tokens.

Supertokens achieve significant compression gains, reducing token fragmentation.

Zero-shot perplexity ratios are consistently lower with Tokenadapt across models.

Abstract

Pretrained language models (LLMs) are often constrained by their fixed tokenization schemes, leading to inefficiencies and performance limitations, particularly for multilingual or specialized applications. This tokenizer lock-in presents significant challenges. standard methods to overcome this often require prohibitive computational resources. Although tokenizer replacement with heuristic initialization aims to reduce this burden, existing methods often require exhaustive residual fine-tuning and still may not fully preserve semantic nuances or adequately address the underlying compression inefficiencies. Our framework introduces two innovations: first, Tokenadapt, a model-agnostic tokenizer transplantation method, and second, novel pre-tokenization learning for multi-word Supertokens to enhance compression and reduce fragmentation. Tokenadapt initializes new unique token embeddings via a hybrid heuristic that combines two methods: a local estimate based on subword decomposition using the old tokenizer, and a global estimate utilizing the top-k semantically similar tokens from the original vocabulary. This methodology aims to preserve semantics while significantly minimizing retraining requirements. Empirical investigations validate both contributions: the transplantation heuristic successfully initializes unique tokens, markedly outperforming conventional baselines and sophisticated methods including Transtokenizer and ReTok, while our Supertokens achieve notable compression gains. Our zero-shot perplexity results demonstrate that the TokenAdapt hybrid initialization consistently yields lower perplexity ratios compared to both ReTok and TransTokenizer baselines across different base models and newly trained target tokenizers. TokenAdapt typically reduced the overall perplexity ratio significantly compared to ReTok, yielding at least a 2-fold improvement in these aggregate scores.