Title: WAON: A Massive Japanese Image-Text Corpus for Enhancing Cultural Adaptation in Contrastive Vision-Language Models

Abstract: The advancement of contrastive vision-language models has been significantly driven by large-scale pretraining initiatives. Emerging research indicates that discarding English-centric caption filters in favor of global datasets effectively enhances multicultural capabilities. This study investigates whether such broad pretraining is adequate for culture-specific comprehension or if supplementary adaptation using native data can further elevate performance beyond the limits of global pretraining alone. To facilitate this inquiry, we introduce WAON, the most extensive publicly accessible native Japanese image-text dataset derived from native Japanese web sources within Common Crawl, comprising roughly 155 million instances. Additionally, we present WAON-Bench, a hand-curated benchmark for Japanese culture covering 374 distinct classes. Our comparative fine-tuning experiments across various Japanese image-text datasets reveal that models adapted via WAON consistently outperform those trained on English-to-Japanese translated data when evaluated on Japanese cultural benchmarks. We have made both the dataset and the accompanying code publicly available.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC