Hugging Face's logo

Datasets:
Omartificial-Intelligence-Space
/
SaudiDialect-Triplet-21

Tasks:
Feature Extraction
Sentence Similarity
Modalities:
Text
Formats:
parquet
Languages:
Arabic
Size:
1K - 10K
Tags:
saudi
Arabic
Triplet
Libraries:
Datasets
pandas
License:
Dataset card Data Studio Files
xet
Community

Access Request – Provide Required Information

Before accessing this model, please complete the form below.

Please provide the required information to access this model

Log in or Sign Up to review the conditions and access this dataset content.

📂 SaudiDialect-Triplet-21 : Saudi Triplet Dataset (SABER Training Data)

🧩 Dataset Summary

The Saudi Triplet Dataset is a high-quality corpus of 2,964 sentence triplets (Anchor, Positive, Negative) specifically curated to capture the nuances of Saudi Arabic dialects (Najdi, Hijazi, Gulf, etc.).

This dataset was created to fine-tune semantic embedding models such as SABER for tasks like Semantic Search, Retrieval-Augmented Generation (RAG), and Clustering.

It covers 21 distinct domains reflecting real-life Saudi contexts, ranging from Government Services and Finance to Tribal Anthropology and Bedouin Culture.

Team

Special thanks to the exceptional team behind this dataset.

Team

✈️

Travel
Mohammed Alhassan

🍔

Food
Abdulelah Alankari

🛍️

Fashion
Reem Alsuliman

🎓

Education
Joud Aloqla

💼

Work
Nouf Alessa

📱

Tech
Jude Alsubaie

🏋️

Sports
Albara Aseri

🚗

Transport
Wajn Alqahtani

🎬

Entertainment
Muzon Assiri

🏠

Daily Life
Jana Alsuhaibani

💰

Finance
Abdullah Alsalem

🌤️

Weather
Huda Aldawsari

🎉

Events
Shaden Alosaimi

🩺

Medical
Munirah Alsubaie

📢

Social
Mohammed Alziyad

🇸🇦

Culture
Shatha Alotaibi

🌿

Nature
Norah Altwijri

📜

History
Renad Alrifai

🗺️

Geography
Murtada Altarouti

🏛️

Gov
Lama Almutairi

👥

Anthro
Adnan Hawsawi

📊 Dataset Statistics

Statistic Value
Total Triplets 2,964
Total Domains 21
Language Saudi Dialect
Duplicate Anchors 59 (Multi-positive/negative pairings)

📏 Sentence Lengths (Word Count)

The dataset consists primarily of short-to-medium length queries and sentences, typical of search and conversational inputs.

Metric Anchor Positive Negative
Mean 6.42 6.50 5.34
Std Dev 1.85 1.96 1.77
Min 2 2 2
Max 13 15 12

🏙️ Domain Distribution

The dataset is balanced across high-resource topics (Food, Finance) and specific cultural topics (Anthropology, Heritage).

Domain Count
Food 200
Finance & Banking 200
Government Services 200
Medical 200
Sports & Fitness 200
Weather & Seasons 200
Nature & Environment 200
Education 150
Travel 150
History 150
Transportation 109
Entertainment 106
Saudi Anthropology 104
Work & Office 104
Culture & Heritage 102
Shopping & Fashion 100
Technology 100
Communication & Social Media 100
Social Gatherings & Events 100
Daily Life & Household 98
Geography 91

📂 Data Structure

Each row in the dataset represents a training triplet designed for Contrastive Learning (e.g., MNRL).

Column Name Type Description
Anchor String The reference sentence/query in Saudi dialect.
Positive String A sentence semantically similar to the Anchor (paraphrase or answer).
Negative String A sentence semantically dissimilar to the Anchor (different topic or meaning).
Domain String The topic category of the triplet.

📝 Data Samples

Below are real examples from the dataset showing the dialectal variations and domain diversity.

Domain Anchor (Query) Positive (Match) Negative (Mismatch)
Shopping & Fashion أبي فرشه تفك العقد وما تقطع الشعر ابي مشط ما يخرب الشعر وينتفه متى بيوصلني طقم الألماس اللي طلبته؟
Finance & Banking أبغا أفتح محفظة أسهم وأبدأ استثمار بسيط أفكر أبدأ تداول خفيف في الأسهم عن طريق المحفظة ناوي أزور العائلة في القرية الأسبوع الجاي
Culture & Heritage أمس سمعت قصائد عن الشجاعة والفروسية القصايد البدوية معانيها قوية شغلت الغسالة بالغلط
Food السوفليه عندهم فخم السوفليه يذوب بالفم ما وصلت الشحنة
History الوالد كان دايم يذكر مملكة لحيان شفت برنامج يتكلم عن سوق عكاظ طلبي تأخر بالمطعم
Travel وين أحصل على جولات سياحية رخيصة؟ أبغى ألقى عروض سياحية اقتصادية الجو حار وما أقدر أطلع

⚠️ Quality & Integrity

  • Missing Data: There are no missing values in the Anchor, Positive, or Negative columns.
  • Duplicates: There are 59 duplicate anchors. This is intentional in some cases to provide multiple positive pairings for the same query or to enforce separation from different hard negatives.
  • Dialect Intensity: The text ranges from "White Dialect" (understandable by most Arabs) to deep Saudi vernacular (specific to Najd/Hijaz/South).

🛠️ Usage

This dataset is optimized for training sentence transformers using MultipleNegativesRankingLoss.

from datasets import load_dataset

# Load the dataset (Example path)
dataset = load_dataset("Omartificial-Intelligence-Space/Saudi-Triplet-Dataset")

# Print first example
print(dataset['train'][0])
Downloads last month
38

Models trained or fine-tuned on Omartificial-Intelligence-Space/SaudiDialect-Triplet-21

Collection including Omartificial-Intelligence-Space/SaudiDialect-Triplet-21