Pearl Vdr Ar Train Hard Mined

A collection of Arabic culturally-aligned image-text triplets for Visual Document Retrieval (VDR). The dataset was created by the author Omartificial-Intelligence-Space, who replaced metadata-based negatives in a preprocessed source with model-mined hard negatives. It was last updated on 2026-04-22.

Use Cases

• Training contrastive learning models based on the described image-text triplets.
• Benchmarking Arabic visual document retrieval systems based on the hard-negative mining approach.
• Fine-tuning embedding models for culturally-aligned Arabic content based on the dataset's focus.
• Studying the effectiveness of model-mined versus metadata-based hard negatives in retrieval tasks.

Strengths

• Includes hard negatives mined by a specific model (Qwen3-VL-Embedding-2B-Arabic-VDR v2, 3-epoch finetune).
• Derived from a preprocessed source dataset (Pearl-vdr-ar-train-preprocessed), suggesting a structured creation process.
• Focuses on Arabic cultural alignment, which may address a specific niche in multimodal data.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

Derived from Omartificial-Intelligence-Space/Pearl-vdr-ar-train-preprocessed.

Collection Method

Negatives were replaced with the top-4 most similar non-matching images according to a specified embedding model.

Freshness

Last updated 2026-04-22 10:21:13; freshness should be verified.