Prompt templates may explain broad behavior shifts after narrow LLM fine-tuning
An author announcing a new paper suggests that broader behavior changes after narrow fine-tuning may depend on the prompting template rather than what a query means.
TLDR
Why can fine-tuning large language models on a narrow domain shift their behavior more broadly? An author announcing a new paper suggests that this generalization may depend on the prompting template, rather than the meaning of the query.
Combined views
13.2K
2 Sources, first seen ago
