DDPM

该数据集旨在评估数据归因方法，特别是在评估扩散模型性能方面的表现，它关注的是当模型在原始数据子集上进行训练时，预测输出变化的能力。数据集采用了子采样技术，并包含了线性数据建模得分（LDS）等指标，以及在不使用最具影响力数据的情况下重新训练模型的方法，以评估不同的数据归因方法。该数据集的规模涉及在原始数据子集上训练的模型，其任务专注于扩散模型中的数据归因。

This dataset is designed to evaluate data attribution methods, especially their performance in assessing diffusion model capabilities. It focuses on the ability to predict changes in model outputs when the model is trained on subsets of the original dataset. The dataset employs subsampling techniques, incorporates metrics such as Linear Data Modeling Scores (LDS), and offers methods for retraining models without utilizing the most impactful training data to evaluate various data attribution approaches. The scope of this dataset covers models trained on subsets of the original data, with its tasks centered on data attribution in diffusion models.