deepghs/csip_v1

---
task_categories:
- zero-shot-image-classification
tags:
- art
- anime
- style-classification
size_categories:
- 100K<n<1M
language:
- multilingual
license:
- cc-by-4.0
source_datasets:
- original
---

# CSIP v1: Contrastive Anime Style Image Pre-training Dataset

## Summary

The **CSIP v1** dataset represents a **roughly cleaned** version of the Contrastive anime Style Image Pre-training collection, specifically designed for **zero-shot image classification** tasks in the anime art domain. This comprehensive dataset contains diverse images from various **anime artists**, organized to facilitate style recognition and classification models. The dataset has been processed through initial cleaning procedures to remove obvious duplicates and low-quality samples while preserving the rich stylistic diversity that characterizes different anime creators.

This dataset serves as an intermediate step between the raw, unprocessed collection and the meticulously curated evaluation set, offering researchers and developers a **balanced compromise** between data volume and quality. The images are distributed across multiple zip archives (p0-p8) for efficient downloading and processing, making it suitable for large-scale pre-training applications where both quantity and reasonable quality are essential considerations.

The **CSIP v1** dataset is particularly valuable for **contrastive learning** approaches, where the stylistic differences between artists can be leveraged to train models that understand and recognize artistic signatures. With its size category of 100K to 1M samples, this dataset provides sufficient scale for training robust vision models while maintaining enough quality control to ensure meaningful learning outcomes. The dataset's organization supports various computer vision tasks beyond zero-shot classification, including style transfer, artist identification, and content-based image retrieval in the anime domain.

## Dataset Structure

The dataset is split into 9 zip archives for convenient downloading and processing:

- `csip_v1_p0.zip` to `csip_v1_p8.zip`

Each archive contains a portion of the cleaned anime style images organized by artist categories.

## Related Datasets

This repository is part of the CSIP dataset series:

- **Raw version**: [deepghs/csip](https://huggingface.co/datasets/deepghs/csip) - The original unprocessed collection
- **Cleaned version**: [deepghs/csip_v1](https://huggingface.co/datasets/deepghs/csip_v1) - This roughly cleaned version
- **Evaluation version**: [deepghs/csip_eval](https://huggingface.co/datasets/deepghs/csip_eval) - Human-picked subset for evaluation

## Usage

The dataset can be downloaded and used for various computer vision tasks, particularly for anime style classification and recognition. The zip archives can be extracted to access the image files organized by artist styles.

## Citation

```bibtex
@misc{csip_v1,
title = {CSIP v1: Contrastive Anime Style Image Pre-training Dataset},
author = {deepghs},
howpublished = {\url{https://huggingface.co/datasets/deepghs/csip_v1}},
year = {2023},
note = {Roughly cleaned version of anime style images for zero-shot classification},
abstract = {The CSIP v1 dataset represents a roughly cleaned version of the Contrastive anime Style Image Pre-training collection, specifically designed for zero-shot image classification tasks in the anime art domain. This comprehensive dataset contains diverse images from various anime artists, organized to facilitate style recognition and classification models. The dataset has been processed through initial cleaning procedures to remove obvious duplicates and low-quality samples while preserving the rich stylistic diversity that characterizes different anime creators.},
keywords = {anime, style-classification, zero-shot-learning, computer-vision}
}
```

原始信息汇总

数据集概述

任务类别

• zero-shot-image-classification

标签

• art

数据集大小

• 100K<n<1M