shiima/kurdish-bidirectional-translation-finall

Kurdish Bidirectional Translation Dataset

Overview

This dataset contains bidirectional translation pairs for Kurdish (Central Kurdish - Sorani script) with English, Arabic, and Persian.

Each sample includes:

• translation: Forward translation (Source → Kurdish)
• back_translation: Reverse translation (Kurdish → Source)

Dataset Details

• Total Samples: 627,103
• Languages: Kurdish (ckb), English (en), Arabic (ar), Persian (fa)
• Format: JSON conversations with instructions

Structure

{
  "id": "xp3x_english_123",
  "translation": '[{"user":"تکایە ئەم دەقە لە ئینگلیزییەوە وەربگێڕە بۆ کوردی\n\n[English text]"},{"assistant":"[Kurdish text]"}]',
  "back_translation": '[{"user":"تکایە ئەم دەقە لە کوردییەوە وەربگێڕە بۆ ئینگلیزی\n\n[Kurdish text]"},{"assistant":"[English text]"}]',
  "source_dataset": "xP3x_english"
}

Preprocessing

All Kurdish text preprocessed with asosoft:

• Normalize
• NormalizePunctuations
• TrimLine

Usage

from datasets import load_dataset
import json

# Load dataset
ds = load_dataset("shiima/kurdish-bidirectional-translation-finall")

# Get a sample
sample = ds['train'][0]

# Parse translation
translation = json.loads(sample['translation'])
print(f"User: {translation[0]['user']}")
print(f"Assistant: {translation[1]['assistant']}")

Source Datasets

• HuggingFaceFW/finetranslations
• CohereLabs/xP3x

License

Apache 2.0

---

Created: 2026-02-08 Pipeline: Memory-optimized chunked processing