ArabicRAGB: Arabic Retrieval-Augmented Generation Benchmark

Dataset Description

ArabicRAGB is a benchmark dataset for evaluating Retrieval-Augmented Generation (RAG) systems on Arabic language tasks. Each record contains a query-passage pair where the query is grounded in the passage content.

Key Features

• Passage-Grounded Queries: Each query is generated from and answerable by its paired passage
• Multi-Dialect Coverage: MSA, Egyptian, Gulf, Levantine, and Maghrebi Arabic
• Complexity Levels: Simple, Moderate, Complex, and Multi-hop queries
• Unified Format: Query and passage together in each record

Dataset Statistics

Component	Count
Training Records	10,530
Test Records	2,633
Total Records	13,163

Dialect Distribution

Dialect	Count
MSA	4,012
Egyptian	3,999
Gulf	3,783
Levantine	697
Maghrebi	672

Complexity Distribution

Level	Count
Simple	4,092
Moderate	4,030
Complex	4,215
Multi-hop	826

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("HeshamHaroon/ArabicRAGB")

train_data = dataset["train"]
test_data = dataset["test"]

print(f"Train records: {len(train_data)}")
print(f"Test records: {len(test_data)}")

# Example record
example = train_data[0]
print(f"Query: {example['query']}")
print(f"Dialect: {example['query_dialect']}")
print(f"Passage: {example['passage_text'][:200]}...")

Example: RAG Evaluation

from datasets import load_dataset
from sentence_transformers import SentenceTransformer
import numpy as np

# Load dataset
dataset = load_dataset("HeshamHaroon/ArabicRAGB")
test_data = dataset["test"]

# Load multilingual model
model = SentenceTransformer('sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2')

# Build passage index from unique passages
passages = list({r["passage_id"]: r["passage_text"] for r in test_data}.values())
passage_embeddings = model.encode(passages, show_progress_bar=True)

# Evaluate retrieval
def retrieve(query, top_k=5):
    query_emb = model.encode(query)
    scores = np.dot(passage_embeddings, query_emb)
    top_indices = np.argsort(scores)[-top_k:][::-1]
    return [passages[i] for i in top_indices]

# Test with a query
results = retrieve(test_data[0]["query"])
print(f"Query: {test_data[0]['query']}")
print(f"Top result: {results[0][:200]}...")

Data Format

Each record contains a query paired with its source passage:

{
  "id": "r_abc123def456",
  "query": "ما هي شروط الحصول على تأشيرة دخول السعودية؟",
  "query_dialect": "msa",
  "query_complexity": "simple",
  "passage_id": "p_xyz789",
  "passage_title": "تأشيرة السعودية",
  "passage_text": "للحصول على تأشيرة دخول المملكة العربية السعودية...",
  "source_url": "https://example.com/visa",
  "source_category": "government"
}

Field Descriptions

Field	Description
id	Unique record identifier
query	Arabic question answerable from the passage
query_dialect	Dialect: msa, egyptian, gulf, levantine, maghrebi
query_complexity	Complexity: simple, moderate, complex, multi_hop
passage_id	Unique passage identifier
passage_title	Title of the source document
passage_text	Full passage text (query answer is within)
source_url	Original source URL
source_category	Category: wikipedia, government, healthcare, etc.

Topics Covered

The corpus covers diverse topics relevant to Arabic speakers:

Category	Topics
Geography	Arab countries, major cities
History	Islamic history, regional history
Culture	Arabic literature, poetry
Science	AI, computing, technology
Health	Medical conditions, healthcare
Law	Legal systems, rights
Economy	Finance, investments

Evaluation Metrics

Recommended metrics for benchmarking:

Retrieval:

• MRR@k (Mean Reciprocal Rank)
• Recall@k
• NDCG@k

Generation:

• BLEU, ROUGE-L
• BERTScore (multilingual)

Limitations

• Queries are synthetically generated (grounded in passages)
• No human-annotated relevance scores
• Content reflects December 2025 snapshot

Citation

@dataset{arabicragb2025,
  title={ArabicRAGB: Arabic Retrieval-Augmented Generation Benchmark},
  author={Hesham Haroun},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/HeshamHaroon/ArabicRAGB}
}

License

This dataset is released under CC BY-SA 4.0.

---

Built for advancing Arabic NLP research