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

In three linesVocabulary adaptation approach to improve LLM efficiency on specialized domains (legal, medical). Combines tokenizer adaptation with selective pretraining on Llama-3.1-8B and Qwen2.5-7B. Reduces training time by 35-55% and parameters by 37% vs expansion-only methods.

Summary generated by Claude — human-verified