From Where Words Come: Efficient Regularization of Code Tokenizers Through Source Attribution

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chizhov, Pavel, Bogomolov, Egor, Yamshchikov, Ivan P.
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910132059766784
author Chizhov, Pavel
Bogomolov, Egor
Yamshchikov, Ivan P.
author_facet Chizhov, Pavel
Bogomolov, Egor
Yamshchikov, Ivan P.
contents Efficiency and safety of Large Language Models (LLMs), among other factors, rely on the quality of tokenization. A good tokenizer not only improves inference speed and language understanding but also provides extra defense against jailbreak attacks and lowers the risk of hallucinations. In this work, we investigate the efficiency of code tokenization, in particular from the perspective of data source diversity. We demonstrate that code tokenizers are prone to producing unused, and thus under-trained, tokens due to the imbalance in repository and language diversity in the training data, as well as the dominance of source-specific, repetitive tokens that are often unusable in future inference. By modifying the BPE objective and introducing merge skipping, we implement different techniques under the name Source-Attributed BPE (SA-BPE) to regularize BPE training and minimize overfitting, thereby substantially reducing the number of under-trained tokens while maintaining the same inference procedure as with regular BPE. This provides an effective tool suitable for production use.
format Preprint
id arxiv_https___arxiv_org_abs_2604_14053
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle From Where Words Come: Efficient Regularization of Code Tokenizers Through Source Attribution
Chizhov, Pavel
Bogomolov, Egor
Yamshchikov, Ivan P.
Computation and Language
Efficiency and safety of Large Language Models (LLMs), among other factors, rely on the quality of tokenization. A good tokenizer not only improves inference speed and language understanding but also provides extra defense against jailbreak attacks and lowers the risk of hallucinations. In this work, we investigate the efficiency of code tokenization, in particular from the perspective of data source diversity. We demonstrate that code tokenizers are prone to producing unused, and thus under-trained, tokens due to the imbalance in repository and language diversity in the training data, as well as the dominance of source-specific, repetitive tokens that are often unusable in future inference. By modifying the BPE objective and introducing merge skipping, we implement different techniques under the name Source-Attributed BPE (SA-BPE) to regularize BPE training and minimize overfitting, thereby substantially reducing the number of under-trained tokens while maintaining the same inference procedure as with regular BPE. This provides an effective tool suitable for production use.
title From Where Words Come: Efficient Regularization of Code Tokenizers Through Source Attribution
topic Computation and Language
url https://arxiv.org/abs/2604.14053