Distilling Token-Trained Models into Byte-Level Models

TL;DR

This paper introduces an efficient method to convert existing token-trained language models into byte-level models through a two-stage distillation process, reducing training costs while maintaining performance.

Contribution

It presents a novel two-stage curriculum for distilling token-trained models into byte-level models, enabling cost-effective byte-level language modeling.

Findings

Distilled BLMs retain most of the teacher models' performance.

The approach requires only approximately 125 billion bytes for training.

Validated across multiple model families like Llama, Qwen, and OLMo.

Abstract

Byte Language Models (BLMs) have emerged as a promising direction for scaling language models beyond tokenization. However, existing BLMs typically require training from scratch on trillions of bytes, making them prohibitively expensive. In this paper, we propose an efficient distillation recipe that converts existing token-trained LLMs into BLMs while retaining comparable capabilities. Our recipe follows a two-stage curriculum: (1) Progressive Knowledge Distillation, which aligns byte-level representations with the embeddings of the token-trained teacher model; and (2) Byte-Level Supervised Fine-Tuning, which enables end-to-end generation entirely in the byte space. We validate our approach across multiple model families, including Llama, Qwen, and OLMo, and demonstrate that the distilled BLMs retain most of the teacher models' performance using only approximately 125B bytes.