In the quest to facilitate the deep intelligence of Large Language Models (LLMs) accessible in final-end user-bot interactions, the art of prompt crafting emerges as a critical yet complex task for the average user. Contrast to previous model-oriented yet instruction-agnostic Automatic Prompt Optimization methodologies, yielding polished results for predefined target models while suffering rapid degradation with out-of-box models, we present Free-form Instruction-oriented Prompt Optimization (FIPO). This approach is supported by our large-scale prompt preference dataset and employs a modular fine-tuning schema. The FIPO schema reimagines the optimization process into manageable modules, anchored by a meta prompt that dynamically adapts content. This allows for the flexible integration of the raw task instruction, the optional instruction response, and the optional ground truth to produce finely optimized task prompts. The FIPO preference dataset is meticulously constructed using the optimal and suboptimal LLMs, undergoing rigorous cross-verification by human experts and analytical models. Applying the insights from the data with Tulu2 models and fine-tuning strategies, we validate the efficacy of FIPO schema across five public benchmarks. Codes, data and scripts are here: https://github.com/LuJunru/FIPO_Project.

大规模语言模型中的自由形式指导导向的提示优化（FIPO）通过模块化微调模式，灵活集成原始任务指令、可选指令响应和可选的目标以生成优化的任务提示。通过使用 Tulu2 模型和微调策略验证 FIPO 模式在五个公共基线上的有效性。

FIPO：使用偏好数据集和模块化微调架构的自由指令导向提示优化