IFLLM
收藏资源简介:
IFLLM数据集是由马萨诸塞大学阿默斯特分校和约克大学的研究团队构建的首个多模态隐式反馈数据集,旨在通过记录用户与大型语言模型交互过程中的眼动轨迹和鼠标移动数据,以解决显式反馈稀缺且成本高昂的问题。该数据集包含1336条多轮问答交互,数据来源于59名亚马逊土耳其机器人工作者,覆盖数百个维基百科主题,通过定制化网站收集了高精度的眼动与鼠标轨迹特征,并附有点对点和成对偏好标注。数据集创建过程涉及基于网络摄像头的眼动校准、多轮对话设计以及严格的质量控制流程,确保了数据的真实性与可靠性。其核心应用领域在于提升LLM对齐效果,通过隐式行为信号训练奖励模型,显著增强偏好预测准确性和响应质量,为下一代个性化对齐方法奠定基础。
The IFLLM dataset is the first multimodal implicit feedback dataset constructed by research teams from the University of Massachusetts Amherst and York University. It aims to address the scarcity and high cost of explicit feedback by recording eye movement trajectories and mouse movement data during user interactions with large language models. This dataset contains 1,336 multi-turn question-answering interactions sourced from 59 Amazon Mechanical Turk workers, covering hundreds of Wikipedia topics. High-precision eye movement and mouse trajectory features were collected via a customized website, paired with point-wise and pairwise preference annotations. The dataset creation process involves webcam-based eye movement calibration, multi-turn dialogue design, and strict quality control procedures, ensuring the data's authenticity and reliability. Its core application lies in improving LLM alignment: training reward models using implicit behavioral signals, which significantly enhances preference prediction accuracy and response quality, laying a foundation for next-generation personalized alignment methods.




