WAON: A Large-Scale, High-Quality Japanese Image-Text Pair Dataset

Introduction

Vision-Language Models (VLMs), which process two modalities (images and text), have been attracting attention in recent years. Datasets of images paired with their corresponding captions (image-text pair datasets) play an important role in VLM development. For example, CLIP, which trains its model with contrastive learning, uses an image-text pair dataset of about 400M pairs. Given the importance of image-text pair datasets, large-scale ones such as LAION-5B are being actively built.

However, existing datasets are skewed toward languages such as English and Chinese, and Japanese image-text pair datasets are insufficient in both scale and quality.

We therefore built WAON (Web-scale image text Aligned Open Nihongo), a large-scale, high-quality Japanese image-text pair dataset.

Overview of WAON

WAON is a Japanese image-text pair dataset consisting of about 155M examples. WAON is built from Japanese web pages collected from Common Crawl, using the following six snapshots: {2025-18, 2025-08, 2024-51, 2024-42, 2024-33, 2024-26}

The data construction pipeline is shown below. To ensure high quality, we carefully performed filtering and deduplication. The WAON construction pipeline

The WAON construction pipeline. The numbers in parentheses are the number of examples remaining after each step for the 2025-18 snapshot.

Evaluating WAON

One way to evaluate the quality of a training dataset itself is to train a model on it and measure the performance of the resulting model. We evaluated the usefulness of WAON by fine-tuning the multilingual CLIP model google/siglip2-base-patch16-256 on WAON and evaluating it on several benchmarks, including Japanese cultural image classification datasets.

WAON-Bench

The Recruit dataset is an existing Japanese cultural image classification dataset, but it has problems: it contains mislabeled examples, and 101 of its 161 classes belong to the food category, so its category distribution is heavily skewed.

We therefore built WAON-Bench, a new benchmark for Japanese cultural understanding. WAON-Bench consists of 1,870 images in total (8 categories, 374 classes, 5 images per class), and each image is annotated with the correct class label.

Overview of WAON-Bench

Overview of WAON-Bench.

Results

The figure below compares Top-1 Accuracy on WAON-Bench when fine-tuning on WAON and on the Japanese subset of ReLAION. Using WAON improves performance efficiently.

Comparison of Top-1 Accuracy on WAON-Bench

Comparison of Top-1 Accuracy on WAON-Bench.

The table below shows the results of each model on several benchmarks. The model fine-tuned on WAON achieves SoTA performance on the Recruit dataset and WAON-Bench. WAON also outperforms the Japanese subset of ReLAION on all benchmarks.

Model	Params	XM3600	ImageNet	Recruit	WAON-Bench	Avg
siglip2-base-patch16-256 (fine-tuned on WAON)	375M	73.75	49.61	83.14	94.97	75.37
siglip2-base-patch16-256 (fine-tuned on ReLAION)	375M	72.39	47.38	81.65	92.99	73.60
siglip2-base-patch16-256	375M	38.28	48.12	76.98	87.81	62.80
clip-japanese-base	196M	78.00	48.90	81.65	90.05	74.65
siglip-base-patch16-256-mult	371M	43.22	53.26	75.10	89.25	65.21
Japanese Stable CLIP ViT-L-16	414M	66.03	55.97	71.29	82.03	68.83
LAION-CLIP-ViT-H-14	1193M	72.64	47.67	70.62	85.88	69.20

These results show that WAON is a high-quality Japanese image-text pair dataset that efficiently improves model performance.

Conclusion

In this post, we introduced WAON, a large-scale, high-quality Japanese image-text pair dataset. Our experiments showed that using WAON efficiently improves a model’s understanding of Japanese culture. For details, see our paper, “WAON: Large-Scale Japanese Image-Text Dataset for Cultural Adaptation in Contrastive Vision-Language Models”.

If you are developing a vision-language model, give WAON a try!