deepghs/csip_eval

---
task_categories:
- zero-shot-image-classification
tags:
- art
- anime
- evaluation
- style-classification
size_categories:
- 1K<n<10K
language:
- multilingual
license: cc-by-4.0
---

# CSIP Evaluation Dataset

## Summary

The **CSIP Evaluation Dataset** is a **human-curated** collection of anime-style images specifically designed for evaluating **zero-shot image classification** models trained on artistic styles. This dataset represents the **highest quality subset** from the Contrastive anime Style Image Pre-Training (CSIP) project, containing carefully selected images that showcase distinct artistic characteristics from various anime artists. The dataset serves as a **benchmark evaluation tool** for assessing model performance on style recognition tasks in the anime art domain.

This evaluation dataset is distinguished by its **human-picked selection process**, where each image has been manually reviewed and approved for quality and stylistic consistency. The curation ensures that the dataset contains **representative examples** of different artistic styles while maintaining high visual quality standards. Unlike automated cleaning methods, this human selection process guarantees that the evaluation samples accurately reflect the intended artistic characteristics, making it particularly valuable for **fine-grained style analysis** and model validation.

The dataset is optimized for **evaluation purposes** rather than training, providing a reliable test bed for models that need to distinguish between different anime art styles. With its carefully balanced composition and quality-controlled samples, it enables **meaningful performance comparisons** across different model architectures and training approaches. The **contrastive learning framework** that this dataset supports allows models to learn discriminative features that capture subtle stylistic differences between artists.

**Keywords**: **human-curated**, **zero-shot classification**, **anime styles**, **evaluation benchmark**, **contrastive learning**

## Dataset Structure

This repository contains two versions of the evaluation dataset:

- `eval_dataset_v0.zip` - Full evaluation dataset
- `eval_dataset_v0_tiny.zip` - Smaller subset for quick testing

## Related Datasets

This evaluation dataset is part of the CSIP (Contrastive anime Style Image Pre-Training) project family:

- **Raw Dataset**: [deepghs/csip](https://huggingface.co/datasets/deepghs/csip) - Original uncleaned dataset
- **Cleaned Dataset**: [deepghs/csip_v1](https://huggingface.co/datasets/deepghs/csip_v1) - Roughly cleaned version for training
- **Evaluation Dataset**: [deepghs/csip_eval](https://huggingface.co/datasets/deepghs/csip_eval) - Current human-picked version (smaller, optimized for evaluation)

## Usage

The dataset is provided in ZIP format for direct download and use. Extract the contents to access the organized image files categorized by artist styles.

```bash
# Download and extract the dataset
unzip eval_dataset_v0.zip
# or for the tiny version
unzip eval_dataset_v0_tiny.zip
```

## Citation

```bibtex
@misc{csip_eval_dataset,
title = {CSIP Evaluation Dataset: Human-Curated Anime Style Images for Zero-Shot Classification},
author = {deepghs},
howpublished = {\url{https://huggingface.co/datasets/deepghs/csip_eval}},
year = {2023},
note = {Human-picked evaluation dataset for anime style classification containing high-quality images from various artists},
abstract = {The CSIP Evaluation Dataset is a human-curated collection of anime-style images specifically designed for evaluating zero-shot image classification models trained on artistic styles. This dataset represents the highest quality subset from the Contrastive anime Style Image Pre-Training (CSIP) project, containing carefully selected images that showcase distinct artistic characteristics from various anime artists. The dataset serves as a benchmark evaluation tool for assessing model performance on style recognition tasks in the anime art domain.},
keywords = {human-curated, zero-shot classification, anime styles, evaluation benchmark, contrastive learning}
}
```

原始信息汇总

数据集概述

任务类别

• zero-shot-image-classification

标签

• art

数据规模

• 1K<n<10K

数据来源

• 人工筛选