Learning Faster with Better Tokens: Parameter-Efficient Vocabulary Adaptation for Specialized Text Summarization

In three linesParameter-efficient vocabulary adaptation method to improve LLM tokenization on specialized domains (legal, medical). Tested on Llama-3.1-8B and Qwen2.5-7B: reduces training time by 35-55% vs continual pretraining, decreases parameters by 37% vs expansion-only, improves summary quality through domain-specific tokens.

Summary generated by Claude — human-verified