PlanE framework optimizes data and tuning for extractive LLMs
New planning framework reduces annotation costs and boosts LLM task efficiency.
PlanE is a new meta-planning framework designed to enhance extractive-based Large Language Models (LLMs). It addresses the high annotation cost of instruction-tuning datasets and lack of optimization methods for specific tasks. The framework consists of three components: data decomposition (breaking down complex tasks), instruction tuning (fine-tuning on task-specific instructions), and prompt inference (crafting effective prompts). Together, these streamline the construction of LLMs for extractive tasks like question answering or information retrieval.
A key innovation is the Data-Tuning-Inference (DTI) planner, which automatically selects the optimal base-LLM and combination of data, tuning, and inference strategies for a given dataset. This improves construction efficiency and generalizes across different optimization objectives. Experiments validate PlanE's effectiveness both across datasets using the same base-LLM and on the same dataset using different base-LLMs. The open-source code allows practitioners to adapt the framework to their own needs.
- PlanE includes data decomposition, instruction tuning, and prompt inference for extractive LLMs.
- DTI planner automatically selects optimal base-LLM and DTI combinations per dataset.
- Validated across multiple datasets and base-LLMs; code publicly available.
Why It Matters
Reduces cost and complexity of building specialized LLMs for extraction tasks, enabling faster deployment.