ProfLLM: A framework for adapting offline large language models to few-shot expert knowledge

Large language models perform well at common field, but are much less effective in niche academic fields .The cause of this problem is that large language models lack the ability to handle few-shot expert knowledge. To address this issue, we purpose ProfLLM, a framework for adapting offline large language models to few-shot expert. ProfLLM provides expert knowledge processing capabilities for large language models through vector databases and prompt engineering, serializes the Q-A pairs of unstructured specialized knowledge through associated databases and completes parameter persistence by embedding fine-tuning, large language model fine-tuning and transfer learning. Experiments show that while ProfLLM performs on par with current mainstream fine-tuning schemes in handling specialized tasks with few-shot expert knowledge, it significantly outperforms fine-tuning schemes in handling general tasks. ProfLLM can handling few-shot expert knowledge while preserving the general capability of the large language model.