Viet OCR VQA Flash2: 137,000 Vietnamese Images with 822,000 Q&A Pairs

137,000 images containing Vietnamese text paired with 822,679 synthetic visual question-answering pairs generated by Gemini 1.5 Flash. Created by 5CD-AI and updated in February 2026, this collection focuses on Vietnamese OCR and scene understanding.

Use Cases

• Training VQA models to answer questions about Vietnamese text within images
• Fine-tuning image-to-text models using the detailed image descriptions
• Evaluating OCR performance on Vietnamese characters across diverse visual contexts

Strengths

• 137,000 images with Vietnamese textual content
• 822,679 individual question-answer pairs
• Labels generated by Gemini 1.5 Flash, a leading VQA model

Limitations

• Synthetic data bias from the Gemini 1.5 Flash model
• Potential for model hallucinations in self-generated Q&A pairs
• Lack of human-verified ground truth for the 822k questions

Provenance

Source

5CD-AI (Arxiv:2408.12480)

Collection Method

Synthetic generation using Gemini 1.5 Flash

Freshness

Last updated February 2026

Geography

Vietnam