ESC-Pro
收藏资源简介:
ESC-Pro数据集是由哈尔滨工业大学的研究团队构建的高质量情感支持对话偏好数据集。该数据集通过蒙特卡洛树搜索(MCTS)扩展已有的标准情感支持对话,生成具有策略偏好的对话树,并从中提取高质量策略偏好对。ESC-Pro不仅包含了优选策略示例,还包含了非优选策略,为模型提供了丰富的训练信号,帮助学习细微的策略权衡,提高适应性决策能力。该数据集适用于优化大型语言模型在情感支持对话中的策略选择准确性和适应性。
ESC-Pro Dataset is a high-quality emotional support conversation preference dataset constructed by the research team from Harbin Institute of Technology. This dataset expands existing standard emotional support conversations through Monte Carlo Tree Search (MCTS) to generate dialogue trees with strategic preferences, and extracts high-quality strategic preference pairs from them. ESC-Pro not only includes examples of preferred strategies but also non-preferred strategies, providing rich training signals for models to learn subtle strategic trade-offs and improve their adaptive decision-making capabilities. This dataset is applicable to optimizing the strategy selection accuracy and adaptability of large language models in emotional support conversations.

- 1Chain of Strategy Optimization Makes Large Language Models Better Emotional Supporter哈尔滨工业大学 · 2025年



