Source-linked AI summary
DanQing: An Up-to-Date Large-Scale Chinese Vision-Language Pre-training Dataset
Hengyu Shen, Tiancheng Gu, Bin Qin, Lan Wu, Yuling Wu, Shuo Tan, Zelong Sun, Jun Wang, Nan Wu, Xiang An, Weidong Cai, Ziyong Feng, Kaicheng Yang
TL;DR
Chinese vision-language pre-training lacks recent, large-scale, high-quality open image-text data. DanQing builds such a dataset with systematic filtering and shows consistent advantages over existing Chinese datasets across downstream evaluations and scaling analyses.
Problem
Chinese vision-language pre-training is constrained by scarce, aging, and unreliable large-scale image-text resources.
Method
DanQing constructs a nearly 100M-pair Chinese dataset from recent web data using filtering for text, visual quality, redundancy, safety, and cross-modal alignment.
Results
DanQing consistently outperforms existing Chinese datasets across downstream benchmarks, including a 50.1% average score versus 49.5% for Chinese-centric multimodal tasks.
Takeaways & Limitations
DanQing provides an up-to-date open resource with stronger scaling capability, higher textual quality, and more balanced visual semantics for Chinese vision-language research.
Abstract
from arXiv · showhide
Vision-Language Pre-training (VLP) models have achieved remarkable success by leveraging large-scale image-text pairs. While English-centric models like CLIP and SigLIP benefit from massive datasets (e.g., LAION-400M), the development of Chinese VLP remains bottlenecked by the lack of high-quality, large-scale open-source data. In this paper, we present DanQing, a large-scale Chinese cross-modal dataset containing 100 million high-quality image-text pairs curated from Common Crawl. To ensure superior data quality, we develop an effective systematic pipeline comprising data source selection, text refinement, visual diversification, and cross-modal cross-batch filtering, thereby effectively mitigating the intrinsic noise prevalent in web data. Notably, DanQing incorporates data from 2024-2025, enabling models to capture contemporary semantic trends and emerging concepts. Extensive experiments via continued pretraining of SigLIP2 models demonstrate that DanQing consistently outperforms existing Chinese datasets across diverse downstream tasks, including zero-shot classification, cross-modal retrieval, and Chinese-centric large multimodal model tasks. Furthermore, in-depth analysis of DanQing reveals that it exhibits a more balanced semantic distribution and superior scaling capability compared to existing datasets. To facilitate further research in Chinese vision-language pre-training, we will open-source the DanQing dataset under the Creative Common CC-BY-NC 4.0 license.
1 Introduction
Vision-language pre-training has benefited from large-scale image-text data, but Chinese VLP remains constrained by limited, aging, and unreliable open resources. DanQing addresses this gap with a large, recent Chinese dataset and extensive downstream evaluation.
- Large-scale image-text data underpins successful vision-language pre-training and broad downstream transfer.
- Chinese VLP resources lag behind English-centric datasets, with the latest cited Chinese dataset introduced over three years earlier.
- Inaccessible image URLs in existing Chinese resources impair training and reproducibility.
- DanQing contributes a filtering pipeline, nearly 100 million Chinese image-text pairs collected after 2024, and evaluation across multiple downstream tasks.
2 Related Work
Related work establishes CLIP-style vision-language pre-training as a data- and architecture-driven paradigm, while emphasizing the continuing scarcity of large-scale, high-quality Chinese image-text datasets.
- CLIP demonstrated strong zero-shot recognition and transfer, motivating subsequent architectural and data-quality improvements.
- The scarcity of large-scale, high-quality Chinese image-text datasets bottlenecks Chinese vision-language pre-training.
- Large-scale datasets such as YFCC100M, LAION400M, and COYO-700M have supported the expansion of vision-language pre-training.
3 DanQing Dataset
DanQing combines SigLIP-based training with multi-stage filtering and analysis of visual, textual, and semantic diversity. Its construction uses recent Common Crawl data and retains broad image, text, and topic coverage.
- Training Objective of DanQing: DanQing trains CLIP-style models with the SigLIP objective, treating image-text alignment as independent sigmoid-based binary classification.
- Training Objective of DanQing: The objective uses scaled similarity sij = (vi · tj)/τ + b and avoids batch coupling compared with CLIP’s cross-entropy formulation.
- Curation of DanQing: Common Crawl data from 2024–2025 is filtered through source reliability, language, textual quality, and safety criteria before further curation.
- Curation of DanQing: Visual diversification removes low-quality images and suppresses perceptual and semantic redundancy using image statistics and ChineseCLIP-L14 embeddings.
- Curation of DanQing: Cross-modal filtering retains Chinese-CLIP-L14 distances in [1.06, 1.24] to improve image-text alignment.
- Statistic of DanQing: DanQing contains 2.2B Chinese words, averages 22 words per sample, and spans varied image resolutions and text lengths.
- Statistic of DanQing: BERTopic analysis of a 10M subset identifies prevalent domains including fashion, technology, cuisine, furnishing, tourism, and sports.
4 Experiments and Results
Continued pretraining with DanQing improves Chinese vision-language performance across zero-shot classification, cross-modal retrieval, and Chinese-centric multimodal tasks. DanQing also achieves the highest reported average score on the LMM evaluation.
- Experimental Setup: DanQing is evaluated by continuing SigLIP2 pretraining for two epochs on 16 × A800 (80G) GPUs.The batch size is 768 × 16, with AdamW optimization and a 1e-5 initial learning rate.
- Zero-shot Classification: 7.6%, 7.8%, and 7.7%: DanQing improves zero-shot classification on SigLIP2 B/32, B/16, and L/16, respectively.Against Wukong, DanQing improves performance by 1.9% across all three backbones.
- Cross-Modal Retrieval: 2.4%&2.4%, 2.1%&2.2%, and 2.7%&2.8%: DanQing improves average retrieval performance over Wukong and Zero across three datasets.The comparisons cover Flickr30K-CN, MSCOCO-CN, and MUGE.
- Chinese-Centric Large Multimodal Model Tasks: 50.1% vs. 49.5%: DanQing produces the highest average score on Chinese-centric LMM downstream benchmarks.The evaluation varies only the SigLIP2-L/16 vision encoder while following the LLaVA-NeXT training pipeline and data configuration.
5 Analysis
Analysis finds DanQing has higher textual semantic density, stronger scaling behavior than Wukong, and support for evaluating understanding of post-2024 internet concepts. The analyses compare dataset text, scaling trajectories, distributions, and emergent-concept recognition.
- Text Quality: DanQing exhibits significantly higher semantic word density than Wukong and Zero in randomly sampled 10M-text comparisons.Semantic density is measured using the proportion of nouns, verbs, and adjectives identified with jieba.
- Data Scaling: 10M, 30M, 60M, and 100M: DanQing is compared with Wukong across these data scales using SigLIP2-B/32.The comparison reports average performance on zero-shot classification and retrieval tasks.
- Data Scaling: DanQing consistently outperforms Wukong across all tested data scales, with larger improvements as scale increases.Wukong’s retrieval performance plateaus beyond 30M, while DanQing provides more effective supervision for large-scale pretraining.
- Dataset Distributions: The analysis compares DanQing and Wukong through clustering and similarity distributions of 10M subsets.These analyses provide dataset-level comparisons of visual-semantic organization and similarity patterns.
6 Conclusion
DanQing is presented as an approximately 100M-pair Chinese image-text dataset built to address scarce high-quality resources for Chinese vision-language pretraining. The paper reports stronger performance, scaling, textual quality, novel-concept understanding, and visual-semantic balance than existing Chinese datasets.
- Conclusion: Approximately 100M pairs: DanQing addresses the scarcity of high-quality Chinese cross-modal resources for vision-language pretraining.The dataset is intended to provide a large-scale foundation for Chinese vision-language research.
- Conclusion: DanQing consistently outperforms existing Chinese datasets across multiple downstream benchmarks after continued SigLIP2 pretraining.The reported analyses also identify stronger scaling capability and improved understanding of novel concepts.
- Conclusion: The paper reports higher textual quality and a more balanced visual semantics distribution for DanQing.These findings come from the paper’s comprehensive dataset analysis.
- Appendices: The appendices provide filtration statistics, DanQing image-text examples, and additional analyses of source domains, topic modeling, and word clouds.These materials are organized in Appendices A–C.
A Statistics of Data Filtration
DanQing’s construction statistics summarize a multi-stage filtering pipeline that reduces approximately 1B raw image-text URLs to nearly 100M high-quality pairs. The process removes about 90% of the original web-data noise.
- Data Filtration: Approximately 1B raw image-text URLs are reduced to nearly 100M high-quality image-text pairs.The resulting dataset occupies approximately 12TB of storage.
- Data Filtration: Four filtering stages comprise data source selection, text refinement, visual diversification, and cross-modal cross-batch filtering.Together, these stages filter out 90% of the original noise.
B Examples in DanQing Dataset
DanQing pairs images with Chinese descriptions across varied domains, illustrating broad thematic coverage for general-purpose vision-language pre-training.
- DanQing examples span natural scenery, historical literature, and automotive technology.
- Table 7 summarizes the statistics and categories of DanQing’s top 40 image source domains.
C.1 Source Domain Distribution
DanQing draws image-text pairs from widely used Chinese platforms spanning e-commerce, news media, and search engines, indicating heterogeneous source coverage.
- Most image-text pairs originate from Chinese platforms and applications including Alibaba, Baidu, and ByteDance.
- The source domains cover categories such as e-commerce, news media, and search engines.
C.2 Topic Modeling
DanQing’s topic distribution includes established categories and additional domains such as tourism, design, education, and agriculture, connecting its content to real-world scenarios.
- Topic examples include tourism, design, education, and agriculture alongside previously identified categories.
- The dataset’s thematic breadth includes diverse visual content aligned with everyday life and real-world scenarios.
C.3 Word Cloud
DanQing’s word distribution highlights contemporary Chinese concepts and diverse daily topics, including terms related to China, games, food, and activities.
- The word cloud is generated from tokenized Chinese text using the jieba segmentation module.
- Frequent terms include “2024”, “中国” (China), “游戏” (Game), “美食” (Food), and “活动” (Activity).
- The distribution reflects newly emerging semantic concepts and diverse daily topics in the dataset.
- Figures 10 and 11 provide visualizations of DanQing topics and its 10M-subset text distribution.