Glow-AI/WaterDrum-TOFU

WaterDrum数据集是一个用于评估遗忘效果和实用性的遗忘基准。该数据集基于TOFU语料库，包含基于原始TOFU数据集的水印和未水印问答数据集。它包括六个子集，代表不同的遗忘场景。每个子集都包含完整、保留、遗忘、精确副本和语义副本的划分。README中还介绍了水印细节和加载数据集以及评估水印强度的方法。该数据集适用于各种遗忘算法，并且可以轻松适应其他模型。

WaterDrum is a benchmark for evaluating the effectiveness and practicality of unlearning. The dataset is based on the TOFU corpus and includes both watermarked and unwatermarked question-answering datasets. It contains six subsets representing different unlearning scenarios, each with full, retain, forget, exact_duplicate, and semantic_duplicate splits. The README provides details on watermarking, how to load the dataset, and methods to evaluate watermark strength. The dataset is suitable for various unlearning algorithms and can be easily adapted to other models.