---
title: MMR-AD
canonical_url: "https://www.modelscope.cn/datasets/xcyao00/MMR-AD"
md_url: "https://www.modelscope.cn/datasets/xcyao00/MMR-AD.md"
repository: xcyao00/MMR-AD
last_updated: 2026-04-13
license: "MIT License"
storage_size: "130 GB"
downloads: 4837
stars: 1
---

# MMR-AD

> MMR-AD - xcyao00 在 ModelScope 开源的数据集。MMR-AD: A Large-Scale Multimodal Dataset for Benchmarking General Anomaly Detection with Multimodal Large Language Models

xcyao00/MMR-AD 是 ModelScope 魔搭社区上的数据集，存储大小 130 GB，采用 MIT License 许可。

- **Repository**: xcyao00/MMR-AD
- **License**: MIT License
- **Storage size**: 130 GB
- **Downloads**: 4837
- **Stars**: 1
- **Last updated**: 2026-04-13

Source: https://www.modelscope.cn/datasets/xcyao00/MMR-AD

---

## MMR-AD: A Large-Scale Multimodal Dataset for General Anomaly Detection

![License](https://img.shields.io/badge/License-MIT-blue.svg)
![CVPR 2026](https://img.shields.io/badge/CVPR-2026-red.svg)

> **Note on Download**: We have placed all the bbox and text annotation files in **ANNOS.zip**. Other zip files named by dataset names contain the original images, so the data volume is relatively large. If you already have image data, you can consider only downloading **ANNOS.zip**.

**MMR-AD** is the current largest-scale **Multi-Modal Reasoning-based** industrial Anomaly Detection dataset, designed to facilitate research on Multimodal Large Language Models (MLLMs) for General Anomaly Detection (GAD).

### 🌟 Key Features

- **Large Scale**: Contains **127,137** high-quality industrial images across **188** product categories.
- **Rich Annotations**: Covers **395** anomaly types with **112,875** annotated anomalous regions (bounding boxes).

> **Note on Bounding Boxes**: All `bbox` annotations are normalized to a **448 resolution**. Specifically, we maintain the original aspect ratio and scale the **shorter edge** of the image to 448 pixels. Coordinates should be interpreted based on this resized dimension.

- **Reasoning-Based Text**: Unlike previous datasets, MMR-AD provides detailed **Chain-of-Thought (CoT)** reasoning texts for each sample, enabling models to learn step-by-step analysis and comparison.
- **Diverse Sources**: Aggregated and cleaned from **14** public industrial AD datasets (e.g., MVTecAD, VisA, RealIAD), ensuring real-world domain shifts.
- **Benchmark Ready**: Structured for both training and evaluating MLLM-based GAD models (e.g., train on subsets, evaluate on unseen categories).

### 📊 Dataset Statistics

| Metric | Value |
| :--- | :--- |
| Total Images | 127,137 |
| Product Categories | 188 |
| Anomaly Types | 395 |
| Annotated Regions | 112,875 |
| Text Tokens | ~26 Million |
| Avg. Tokens/Sample | 208 |



📂 Data Structure

The dataset is organized by sub-datasets to facilitate cross-domain evaluation:

```
MMR-AD/
├── MVTecAD/
├── VisA/
├── RealIAD/
└── ... (14 subdatasets)
```

Each sample includes:
- `reference_image`: Normal reference sample.
- `input_image`: Image to be inspected.
- `reasoning_text`: Detailed CoT analysis (hidden in `<think>` tags).
- `answer`: Final detection result with bounding box coordinates and anomaly label.

### 🚀 Usage

MMR-AD is ideal for:

1. **Post-training MLLMs**: Fine-tuning models like Qwen-VL, LLaVA, or InternVL for industrial inspection.

2. **General AD Research**: Evaluating model generalization on unseen product categories.

3. **Reasoning Studies**: Analyzing the impact of Chain-of-Thought on visual anomaly detection.

### 📄 Citation

If you use MMR-AD in your research, please cite our paper:

```
@inproceedings{yao2026mmrad,
  title={MMR-AD: A Large-Scale Multimodal Dataset for Benchmarking General Anomaly Detection with Multimodal Large Language Models},
  author={Yao, Xincheng and Qian, Zefeng and Shi, Chao and Song, Jiayang and Zhang, Chongyang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}
```