ADPL
收藏资源简介:
ADPL数据集是一个针对工业异常推理的偏好优化数据集,包含28,415个问答实例,每个实例包括一对测试-参考图像以及由领域专家排名的四个候选答案。该数据集旨在支持对异常推理模型的质量进行综合评估,通过多尺度逻辑评估框架(MLE)衡量模型在识别精度、定位精度和推理一致性三个关键维度上的性能。
The ADPL dataset is a preference optimization dataset for industrial anomaly reasoning. It contains 28,415 question-answer instances, each consisting of a pair of test-reference images and four candidate answers ranked by domain experts. This dataset aims to support comprehensive quality evaluation of anomaly reasoning models, measuring their performance across three key dimensions: recognition accuracy, localization accuracy, and reasoning consistency, via the Multi-scale Logical Evaluation (MLE) framework.
SAGE数据集概述
数据集基本信息
- 数据集名称: AD-PL (Anomaly Detection and Reasoning Dataset)
- 相关论文: "SAGE: A Visual Language Model for Anomaly Detection via Fact Enhancement and Entropy-aware Alignment" (ACM Multimedia 2025)
- 官方仓库: https://github.com/amoreZgx1n/SAGE
数据集下载
- AD-PL下载地址: https://pan.baidu.com/s/1Jr68D6ysgdEFgOB0UTZJdw?pwd=b2nd
- 其他相关数据集:
- MANTA: https://grainnet.github.io/MANTA
- MMAD: https://github.com/jam-cc/MMAD
数据集用途
- 主要用途: 异常检测与推理
- 适用场景: 视觉语言模型的训练与测试
相关资源
- 预训练模型: InternVL2-8B (https://huggingface.co/OpenGVLab/InternVL2-8B)
- 许可证: Apache-2.0 License
引用信息
bibtex @misc{zang2025sagevisuallanguagemodel, title={SAGE: A Visual Language Model for Anomaly Detection via Fact Enhancement and Entropy-aware Alignment}, author={Guoxin Zang and Xue Li and Donglin Di and Lanshun Nie and Dechen Zhan and Yang Song and Lei Fan}, year={2025}, eprint={2507.07939}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2507.07939}, }

- 1SAGE: A Visual Language Model for Anomaly Detection via Fact Enhancement and Entropy-aware Alignment哈尔滨工业大学 · 2025年



