WAON: Large-Scale and High-Quality Japanese Image-Text Pair Dataset for Vision-Language Models

TL;DR

This paper tackles the scarcity of large-scale, high-quality Japanese image-text datasets for Vision-Language Models. It presents WAON, a ~155M Japanese image-text pair dataset built from Common Crawl using a rigorous multi-stage pipeline, including language filtering, deduplication, image-quality checks, NSFW removal, and SigLIP-based alignment filtering. To evaluate WAON, the authors create WAON-Bench, a manually curated 374-class Japanese cultural image classification benchmark, and compare WAON with the Japanese subset of ReLAION by fine-tuning SigLIP2; results show WAON yields higher data quality and training efficiency, achieving state-of-the-art performance on several Japanese cultural benchmarks. The practical impact is significant: WAON enables better cultural understanding in VLMs and the pipeline is adaptable to other languages, with WAON, the model, and code released publicly for researchers and developers. Overall, the work advances Japanese VLM capabilities by combining scalable data collection with careful quality control and targeted benchmarking.

Abstract

Large-scale and high-quality image-text pair datasets play an important role in developing high-performing Vision-Language Models (VLMs). In this work, we introduce WAON, a large-scale and high-quality Japanese image-text pair dataset containing approximately 155 million examples, collected from Common Crawl. Our dataset construction pipeline employs various techniques, including filtering and deduplication, which have been shown to be effective in previous studies. To evaluate its effectiveness, we also construct WAON-Bench, a manually curated benchmark for Japanese cultural image classification, consisting of 374 classes. To assess the effectiveness of our dataset, we conduct experiments using both WAON and the Japanese subset of ReLAION, one of the most widely used vision-language datasets. We fine-tune SigLIP2, a strong multilingual model, on both datasets. The results demonstrate that WAON enhances model performance on WAON-Bench more efficiently than ReLAION and achieves higher accuracy across all evaluated benchmarks. Furthermore, the model fine-tuned on WAON achieves state-of-the-art performance on several Japanese cultural benchmarks. We release our dataset, model, and code at https://speed1313.github.io/WAON.

Figures (5)

• Figure 1: WAON-Bench Top-1 accuracy over training. Fine-tuning google/siglip2-base-patch16-256 on WAON improves performance on Japanese cultural benchmarks more efficiently than fine-tuning on the ReLAION (Japanese subset).
• Figure 2: Overview of WAON-Bench. WAON-Bench consists of 8 categories with a total of 374 classes, each containing 5 images, resulting in a total of 1,870 images.
• Figure 3: Construction pipeline of WAON. The numbers in parentheses indicate the remaining data count after each processing step (based on the 2025-18 snapshot).
• Figure 4: SigLIP similarity score distribution of image-text pairs from the 2025-18 snapshot. We exclude image-text pairs with similarity scores below 0.1.
• Figure 5: t-SNE map of image embeddings of WAON-Bench.