Dataset · llamaindex
Multilingual dataset for visual document retrieval
llamaindex/vdr-multilingual-train
500k pairs of page images and synthetic queries in five European languages for multimodal retrieval.
What's inside
The dataset contains 500k query-image pairs collected from public web PDF documents. Queries are synthetically generated using VLM models gemini-1.5-pro and Qwen2-VL-72B in English, German, French, Spanish, and Italian. The multimodal embedding model vdr-2b-multi-v1 was trained on these data.
What you can do with it
- Training multilingual multimodal embedding models for PDF and scan search
- Building and evaluating Visual Document Retrieval and RAG systems over visual documents
- Languages
- German, Italian, French, Spanish, English
- Size
- 100K–1M rows
- Downloads this month
- 1.8K
- Likes
- 32
Models trained on it
- llamaindex/vdr-2b-multi-v1Image text → text 15.2K
- athrael-soju/colqwen3.5-4.5B-v3Visual document retrieval 3.4K
- nomic-ai/colnomic-embed-multimodal-7bVisual document retrieval 2.2K
- tsystems/colqwen2.5-3b-multilingual-v1.0Visual document retrieval 1.6K
- Metric-AI/ColQwen2.5-3b-multilingual-v1.0Visual document retrieval 500
- mradermacher/vdr-2b-multi-v1-i1-GGUF 473
- mradermacher/vdr-2b-v1-GGUF 442
- VAGOsolutions/SauerkrautLM-ColQwen3-2b-v0.1Image text → text 438
How to load it
from datasets import load_dataset
ds = load_dataset("llamaindex/vdr-multilingual-train")