Recent studies have assessed the use of state-of-the-art models (e.g., Foundation Models) for code refactoring, but the effectiveness of these pre-trained models remains limited. While model performance can be enhanced through datasets, the lack of high-quality data poses a significant challenge. This paper introduces the MaRV dataset, which contains 693 manually evaluated code pairs extracted from 126 GitHub Java repositories, representing four types of refactoring. Additionally, metadata describing the supposedly refactored elements was collected. Each code pair was manually evaluated by two reviewers from a pool of 40 participants. The MaRV dataset is continuously evolving, with a web-based tool available for evaluating refactoring representations. The primary application of this dataset is to improve the accuracy and reliability of state-of-the-art models in refactoring tasks, such as refactoring candidate identification and refactoring code generation, by providing high-quality data.