Datasetpublic

Omartificial-Intelligence-Space/Pearl-vdr-ar-train-hard-mined

Pearl-vdr-ar-train-hard-mined Arabic culturally-aligned Visual Document Retrieval (VDR) triplets with model-mined hard negatives, derived from Omartificial-Intelligence-Space/Pearl-vdr-ar-train-preprocessed by replacing its metadata-based negatives with the top-4 most similar non-matching images according to Omartificial-Intelligence-Space/Qwen3-VL-Embedding-2B-Arabic-VDR (v2, 3-epoch finetune). Dataset structure Each row contains: Column Type… See the full description on the dataset page: https://huggingface.co/datasets/Omartificial-Intelligence-Space/Pearl-vdr-ar-train-hard-mined.

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

2likes19downloads

Dataset Card

Pearl-vdr-ar-train-hard-mined

Arabic culturally-aligned Visual Document Retrieval (VDR) triplets with model-mined hard negatives, derived from `Omartificial-Intelligence-Space/Pearl-vdr-ar-train-preprocessed` by replacing its metadata-based negatives with the top-4 most similar non-matching images according to `Omartificial-Intelligence-Space/Qwen3-VL-Embedding-2B-Arabic-VDR` (v2, 3-epoch finetune).

Dataset structure

Each row contains:

Column	Type	Description
`query`	string	Arabic text query about a cultural topic
`category`	string	One of 9 topics (Music, Landmarks, Food, Clothes, …)
`country`	string	One of 19 Arab states
`image`	image	Positive document screenshot (what the query should retrieve)
`negative_0`	image	Hard negative — mined top-1 non-matching image by v2 cosine sim
`negative_1`	image	Hard negative — mined top-2 non-matching image by v2 cosine sim
`negative_2`	image	Hard negative — mined top-3 non-matching image by v2 cosine sim
`negative_3`	image	Hard negative — mined top-4 non-matching image by v2 cosine sim

Splits

Config	Rows	File size	Source
`train`	48 002	17.7 GB	Hard-mined negatives

Splits preserve the category-stratified sampling from the parent dataset (9 categories × 19 countries).

Construction

The pipeline:

1.Load parent `Pearl-vdr-ar-train-preprocessed` (train + dev + test).
2.Hash every image across all 5 image columns of all 3 splits by JPEG bytes → 10 299 unique images (the parent dataset has heavy deduplication).
3.Embed the full image pool with Omartificial-Intelligence-Space/Qwen3-VL-Embedding-2B-Arabic-VDR at max_pixels=600×600.
4.Embed every training query.
5.For each query, compute cosine similarity against the entire pool, mask out the positive image, and take the top-4 highest-similarity remaining images as negative_0..3.
6.Emit a new train parquet; copy dev/test through unchanged.

This is a standard hard-negative mining recipe — the same approach that reliably improves English retrieval models. Its effectiveness is tied to how discriminative the miner is in the target language.

How to load

python

from datasets import load_dataset

train = load_dataset("Omartificial-Intelligence-Space/Pearl-vdr-ar-train-hard-mined", "train", split="train")
dev   = load_dataset("Omartificial-Intelligence-Space/Pearl-vdr-ar-train-hard-mined", "dev",   split="train")
test  = load_dataset("Omartificial-Intelligence-Space/Pearl-vdr-ar-train-hard-mined", "test",  split="train")

print(train[0]["query"])
print(train[0]["image"].size, train[0]["negative_0"].size)

Intended use

Primary use case: ablation / reproducibility research on hard-negative mining in low-resource language multimodal retrieval. Also useful for:

—Benchmarking miner quality across different base models (embed pool here, compare against v2's mined top-K).
—Multi-loss training (combine metadata negatives + mined negatives).
—Negative-mining curriculum experiments.

Attribution

—Parent dataset: `Omartificial-Intelligence-Space/Pearl-vdr-ar-train-preprocessed`.
—Original source of images + Q/A pairs: `MohamedRashad/Arabic-VLM-Full-Pearl`.
—Miner model: `Omartificial-Intelligence-Space/Qwen3-VL-Embedding-2B-Arabic-VDR` (v2 commit 2f851aab, before the rolled-back v3 commit).
—Recipe inspired by Tom Aarsen's multimodal Sentence Transformers VDR blog.

Citation

BibTeX

Pearl Dataset

If you use this dataset or the accompanying benchmarks, please cite our paper:

bibtex

@inproceedings{alwajih-etal-2025-pearl,
    title = "Pearl: A Multimodal Culturally-Aware {A}rabic Instruction Dataset",
    author = "Alwajih, Fakhraddin  and
      Magdy, Samar M.  and
      El Mekki, Abdellah  and
      Nacar, Omer  and
      Nafea, Youssef  and
      Abdelfadil, Safaa Taher  and
      Yahya, Abdulfattah Mohammed  and
      Luqman, Hamzah  and
      Almarwani, Nada  and
      Aloufi, Samah  and
      Qawasmeh, Baraah  and
      Atou, Houdaifa  and
      Sibaee, Serry  and
      Alsayadi, Hamzah A.  and
      Al-Dhabyani, Walid  and
      Al-shaibani, Maged S.  and
      El aatar, Aya  and
      Qandos, Nour  and
      Alhamouri, Rahaf  and
      Ahmad, Samar  and
      AL-Ghrawi, Mohammed Anwar  and
      Yacoub, Aminetou  and
      AbuHweidi, Ruwa  and
      Lemin, Vatimetou Mohamed  and
      Abdel-Salam, Reem  and
      Bashiti, Ahlam  and
      Ammar, Adel  and
      Alansari, Aisha  and
      Ashraf, Ahmed  and
      Alturayeif, Nora  and
      Alcoba Inciarte, Alcides  and
      Elmadany, AbdelRahim A.  and
      Tourad, Mohamedou Cheikh  and
      Berrada, Ismail  and
      Jarrar, Mustafa  and
      Shehata, Shady  and
      Abdul-Mageed, Muhammad",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2025",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "[https://aclanthology.org/2025.findings-emnlp.1254/](https://aclanthology.org/2025.findings-emnlp.1254/)",
    pages = "23048--23079",
    ISBN = "979-8-89176-335-7"
}