---
title: WSADBench-Datasets
canonical_url: "https://www.modelscope.cn/datasets/mac4mac/WSADBench-Datasets"
md_url: "https://www.modelscope.cn/datasets/mac4mac/WSADBench-Datasets.md"
repository: mac4mac/WSADBench-Datasets
chinese_name: "WSADBench官方基准数据集"
last_updated: 2026-05-26
license: "Apache License 2.0"
storage_size: "1.1 TB"
downloads: 13171
stars: 0
---

# WSADBench-Datasets

> WSADBench-Datasets - mac4mac 在 ModelScope 开源的数据集。The official dataset repository for WSADBench. Contains multi-modal features (Video/CV) for Weakly-Supervised Anomaly Detection, including OOD updates.

mac4mac/WSADBench-Datasets 是 ModelScope 魔搭社区上的数据集，存储大小 1.1 TB，采用 Apache License 2.0 许可。

- **Repository**: mac4mac/WSADBench-Datasets
- **License**: Apache License 2.0
- **Storage size**: 1.1 TB
- **Downloads**: 13171
- **Stars**: 0
- **Last updated**: 2026-05-26

Source: https://www.modelscope.cn/datasets/mac4mac/WSADBench-Datasets

---

# WSADBench Datasets

This dataset repository provides auxiliary data files, pre-processed features, and mirrored checkpoints used by WSADBench. The files are hosted for research convenience, faster downloads, and benchmark reproducibility.

ModelScope dataset page: https://www.modelscope.cn/datasets/mac4mac/WSADBench-Datasets/files

## File Structure

| Name | Description |
| --- | --- |
| `ckpt/` | Mirrored checkpoints for TabPFN-2.5 and LimiX-16M. |
| `CV_by_I3D/` | Features extracted from four video anomaly detection datasets using I3D preprocessing. |
| `CV_by_MViT_32/` | Features extracted from four video anomaly detection datasets using MViT preprocessing. |
| `CV_by_SlowFast/` | Features extracted from four video anomaly detection datasets using SlowFast preprocessing. |
| `CV_by_SlowFast_R50/` | Features extracted from four video anomaly detection datasets using SlowFast-R50 preprocessing. |
| `CV_by_X3DM/` | Features extracted from four video anomaly detection datasets using X3DM preprocessing. |
| `env/` | Python dependency files required by LimiX. This follows the official LimiX environment setup process. |
| `.gitattributes` | Git LFS and repository attribute configuration. |
| `Classical_bags_inexact.tar.gz` | Tabular datasets preprocessed into MIL bags for inexact-supervision experiments. |
| `CV_by_ResNet18_OOD.tar.gz` | Image datasets preprocessed with ResNet18 features for OOD experiments. |
| `dataset-name-matching.json` | Internal mapping file used by download scripts. Users usually do not need to download it manually. |
| `README.md` | Dataset description, checkpoint notice, license notes, and acknowledgement. |

## Checkpoint Mirror Notice

The pre-trained checkpoints hosted in this repository, including TabPFN-2.5 and LimiX-16M, are mirrors of the officially released weights from their respective authors. We host them here solely for accelerating downloads and ensuring benchmark reproducibility.

All copyrights belong to the original authors. Users must follow the original licenses, model cards, and citation requirements from the official repositories.

## Checkpoint Sources and Licenses

### TabPFN-2.5

- Official files page: https://huggingface.co/Prior-Labs/tabpfn_2_5/tree/main
- Official code repository: https://github.com/PriorLabs/tabPFN
- License: `TABPFN-2.5 Non-Commercial License v1.1`

TabPFN-2.5 is released by Prior Labs under a custom non-commercial license. 

The license allows non-commercial and non-production use, including testing, evaluation, academic research, internal benchmarking, and experimentation. It does not allow commercial use, production deployment, paid services, client deliverables, commercial decision-making, or hosted API/SaaS services without a separate commercial license from Prior Labs.

If you use the mirrored TabPFN-2.5 checkpoints, please follow the original `TABPFN-2.5 Non-Commercial License v1.1`, keep the required attribution notice, and refer to the official Prior Labs model page for the latest license terms.
#### Note
The TABPFN-2.5 Model is licensed by Prior Labs GmbH under the TABPFN-2.5 Non-Commercial License.
Copyright © Prior Labs GmbH 2025.
 THE SERVICES ARE PROVIDED FREE OF CHARGE: COMPANY SHALL NOT BE LIABLE FOR DAMAGES RESULTING FROM SLIGHT NEGLIGENCE. LIABILITY FOR GROSS NEGLIGENCE AND INTENTIONAL MISCONDUCT REMAINS UNAFFECTED.


### LimiX-16M

- Official files page: https://huggingface.co/stable-ai/LimiX-16M/tree/main
- Official code repository: https://github.com/limix-ldm/LimiX
- License note: code under `Apache-2.0`; model weights subject to the LimiX Model License

According to the official LimiX model card, the code is open-sourced under the Apache-2.0 license, while the usage of the LimiX model weights is subject to the Model License. The LimiX weights are fully available for academic research and may be used commercially only after obtaining proper authorization.

If you use the mirrored LimiX-16M checkpoints, please distinguish between the code license and the model-weight license, and refer to the official LimiX model page for the latest authorization requirements.


## Acknowledgement

We sincerely thank Prior Labs for releasing TabPFN-2.5 and the LimiX team for releasing LimiX-16M. Their released models make it possible to evaluate modern tabular foundation models under weakly supervised anomaly detection settings.