ICM-Instruct
收藏资源简介:
ICM-Instruct数据集是一个用于图像内容审核的大规模指令调优数据集,由香港中文大学、华为香港研究中心等机构联合创建。该数据集通过分解人类定义的规则,并利用多阶段提示生成详细的审核解释和问答对,涵盖了多种文化规范和儿童保护标准。数据集的内容包括丰富的图像注释、审核解释和问答对,旨在提升多模态大语言模型在图像内容审核中的分类和解释能力。数据集的创建过程涉及规则分解、图像下载和多阶段提示生成,最终应用于训练ICM-Assistant模型,显著提升了审核分类和解释的准确性。该数据集的应用领域主要是图像内容审核,旨在解决现有审核模型在分类和解释上与人类审核员不一致的问题,提供灵活、可解释且准确的审核结果。
The ICM-Instruct dataset is a large-scale instruction-tuning dataset for image content moderation, jointly created by institutions including The Chinese University of Hong Kong and Huawei Hong Kong Research Center, and other relevant organizations. This dataset decomposes human-defined rules and utilizes multi-stage prompting to generate detailed moderation explanations and question-answer pairs, covering a wide range of cultural norms and child protection standards. The dataset contains rich image annotations, moderation explanations and question-answer pairs, aiming to improve the classification and explanation capabilities of multimodal large language models in image content moderation. The development process of the dataset involves rule decomposition, image downloading and multi-stage prompting generation, and is ultimately applied to train the ICM-Assistant model, which significantly enhances the accuracy of moderation classification and explanation. The main application scenario of this dataset is image content moderation, which aims to address the inconsistency between existing moderation models and human moderators in terms of classification and explanation, and provide flexible, interpretable and accurate moderation results.
ICM-Assistant 数据集概述
数据集结构
- assets: 包含与ICM-Assistant相关的内容。
- data: 包含用于数据生成管道的工具。
- eval: 包含用于性能评估管道的工具。
- inference: 包含用于推理管道的工具。
- training: 包含用于训练管道的工具。
数据集用途
该数据集主要用于支持ICM-Assistant项目的各个流程,包括数据生成、性能评估、推理和训练。每个目录下包含的工具和内容分别对应不同的任务流程,帮助用户完成从数据准备到模型训练和评估的全过程。

- 1ICM-Assistant: Instruction-tuning Multimodal Large Language Models for Rule-based Explainable Image Content Moderation香港中文大学, 华为香港研究中心, 香港科技大学, 华为, 上海交通大学, 新加坡国立大学, 西安电子科技大学, 广州理工学院, ICTT和ISN实验室 · 2024年



