AH-Translit_Benchmark

Arabic → Hindi Transliteration Benchmark Dataset

---

Description

AH-Translit_Bench is a 6,000-sample Arabic-to-Hindi transliteration benchmark dataset, curated for systematic evaluation and comparison of transliteration models across three linguistically distinct domains:

• Quranic Arabic
• Modern Standard Arabic (Daily Use)
• Modern Standard Arabic (Bibliographic)

Each domain contributes 2,000 carefully selected sentence pairs, ensuring balanced, fair, and domain-aware evaluation of cross-script transliteration systems.

This benchmark is designed strictly for testing and reporting results, and is complementary to the full AH-Translit training dataset.

---

Dataset Usage

Designed for benchmarking and evaluating Arabic-to-Hindi phonetic transliteration models under domain shifts and varying sequence-length conditions.

---

Content Type

Text — sentence-level Arabic source text paired with Hindi (Devanagari) phonetic transliteration.

---

File Type

CSV (Comma-Separated Values)

---

Dataset Structure

AH-Translit-Benchmark-Dataset
├── quranic_benchmark_2000.csv
├── msa_dailyuse_benchmark_2000.csv
├── msa_bibliographic_benchmark_2000.csv
├── all_domain_mix_benchmark_6000.csv
└── README.md

---

File Descriptions

• quranic_benchmark_2000.csv 2,000 sentence pairs from Quranic Arabic, characterized by long sequences and rich morphology.

• msa_dailyuse_benchmark_2000.csv 2,000 sentence pairs from daily-use Modern Standard Arabic, representing short and conversational inputs.

• msa_bibliographic_benchmark_2000.csv 2,000 sentence pairs from formal and bibliographic MSA, featuring higher lexical diversity.

• all_domain_mix_benchmark_6000.csv Combined benchmark file containing all 6,000 samples, evenly distributed across domains.

Each CSV contains exactly two columns:

Arabic , Hindi

Hindi text is phonetic transliteration, not translation.

---

Benchmark Scale Summary

Domain	Samples
Quranic	2,000
MSA Daily Use	2,000
MSA Bibliographic	2,000
Total	6,000

This strict balance ensures macro-averaged, unbiased evaluation across domains.

---

How to Use This Dataset

Here is a clean, minimal, copy-paste–ready Python sample that you can put directly in your README.md or use for quick sanity testing after installing the package.

This matches your current AH-Translit-Bench v2.0.0 API.

---

Python Usage Example

1️⃣ Install the Package

pip install AH-Translit-Bench

---

2️⃣ Load Domain-wise Benchmark Datasets

from AH_Translit_Bench import load_dataset, get_available_domains

# List all available domains
domains = get_available_domains()
print("Available domains:", domains)

Expected output

Available domains: ['quranic', 'msa_dailyuse', 'msa_bibliographic', 'all']

---

3️⃣ Load Individual Domains

Quranic Benchmark (2000 samples)

quranic_df = load_dataset("quranic")
print(quranic_df.head())
print("Quranic shape:", quranic_df.shape)

Expected:

Quranic shape: (2000, 2)

---

MSA Daily Use Benchmark (2000 samples)

daily_df = load_dataset("msa_dailyuse")
print(daily_df.head())
print("MSA Daily Use shape:", daily_df.shape)

Expected:

MSA Daily Use shape: (2000, 2)

---

MSA Bibliographic Benchmark (2000 samples)

biblo_df = load_dataset("msa_bibliographic")
print(biblo_df.head())
print("MSA Bibliographic shape:", biblo_df.shape)

Expected:

MSA Bibliographic shape: (2000, 2)

---

4️⃣ Load the Full Mixed Benchmark (All Domains)

all_df = load_dataset("all")
print(all_df.head())
print("All-domain benchmark shape:", all_df.shape)

Expected:

All-domain benchmark shape: (6000, 2)

---

5️⃣ Access Arabic–Hindi Pairs

arabic_text = quranic_df.iloc[0]["Arabic"]
hindi_translit = quranic_df.iloc[0]["Hindi"]

print("Arabic:", arabic_text)
print("Hindi :", hindi_translit)

---

Example Data Snippet

Arabic,Hindi
المطبعة الحيدرية،,"अल-मतबअह अल-हैदरियह,"

---

Version Overview

Version 2.0 (Current)

Version Description: Expanded benchmark release with 2,000 samples per domain, covering Quranic Arabic, MSA Daily Use, and MSA Bibliographic text, along with a combined 6,000-sample mixed-domain benchmark file for standardized evaluation.

---

Important Notes

• This is a benchmark-only dataset
• Not intended for model training
• No overlap with AH-Translit training splits
• Focuses on phonetic fidelity, not semantic translation

---

License

• Code: MIT License
• Dataset: Creative Commons Attribution–NonCommercial 4.0 (CC BY-NC 4.0)

Permitted for research and educational use only with proper attribution.

Full license: https://creativecommons.org/licenses/by-nc/4.0/

---