PlanE Optimizes Extractive LLM Construction and Tuning

A novel planning framework named PlanE (Planning for Extractive-based LLMs) has been introduced to streamline the construction and optimization of extractive large language models. This framework aims to tackle the significant challenges of high annotation costs for instruction-tuning datasets and the absence of effective optimization methods for tailoring LLMs to specific tasks. PlanE integrates three key components: data decomposition, instruction tuning, and prompt inference. Central to PlanE is the Data-Tuning-Inference (DTI) planner, which intelligently selects the most suitable base-LLM and its corresponding DTI combinations for particular datasets. This meta-planning approach is designed to enhance construction efficiency. Experimental results validate PlanE's effectiveness across different datasets using the same base-LLM, and on the same dataset with various base-LLMs. The DTI planner also demonstrates strong generalizability across diverse optimization objectives, making it a versatile tool for developing more efficient and task-specific extractive LLMs.