TajPersParallelCorpusFull

📖 Description

Full version of the Tajik–Persian parallel corpus, combining:

1. TajPersParallelCorpus – 328,253 sentence pairs from news, poetry, prose, and named entities
2. TajPersParallelLexicalCorpus – 175,130 lexical pairs with author attribution

After deduplication, the combined corpus contains 456,307 unique parallel pairs.

📊 Dataset Statistics

Overview

Metric	Value
Total pairs	456,307
Unique Tajik entries	433,345
Unique Persian entries	453,515
Empty Tajik/Persian fields	0 / 0

Length Statistics

Metric	Tajik	Persian
Avg characters	40.0	33.4
Max characters	973	838
Avg tokens (words)	7.0	7.4
Max tokens	157	161

Category Distribution (Top 15)

Category	Count	Percentage
poetry_parts	154,891	33.94%
masnavi	38,972	8.54%
unique_tajik_words	38,151	8.36%
Rumi	31,275	6.85%
shahnameh	24,819	5.44%
prose_parts	23,936	5.25%
Dictionary	23,907	5.24%
Ferdousi	22,290	4.88%
paranames_per	20,101	4.41%
Nezami	14,926	3.27%
words	14,292	3.13%
paranames_loc	9,586	2.10%
Asadi	7,797	1.71%
Saadi	7,266	1.59%
Bidel	4,443	0.97%

Source Composition

The corpus is built from:

• Original HF corpus: 328,253 pairs (72.0% of unique pairs)
• Lexical corpus: 175,130 pairs (38.4% of unique pairs, ~47k duplicates removed)

📁 Data Format

Each record contains three fields:

Field	Type	Description
tajik	string	Tajik text (Cyrillic script)
farsi	string	Persian text (Arabic script)
category	string	Source category (e.g., poetry_parts, bbc, Dictionary, Rumi)

Category Values

• Original categories: poetry_parts, masnavi, unique_tajik_words, shahnameh, prose_parts, paranames_*, bbc, etc.
• Lexical categories: +Saadi format (e.g., Assorted+Dictionary, BBC+Saadi)

🚀 Usage Example

from datasets import load_dataset

# Load the full corpus
dataset = load_dataset("TajikNLPWorld/TajPersParallelCorpusFull")
data = dataset["train"]

# Get first record
print(data[0])

# Filter by category
poetry = data.filter(lambda x: x["category"] == "poetry_parts")
bbc_news = data.filter(lambda x: x["category"] == "bbc")
lexical = data.filter(lambda x: "+" in x["category"])  # lexical entries have "source+author"

print(f"Poetry pairs: {len(poetry)}")
print(f"BBC news pairs: {len(bbc_news)}")
print(f"Lexical pairs: {len(lexical)}")

🔬 Applications

• Machine Translation – Tajik ↔ Persian
• Cross-lingual Retrieval – search in one language, retrieve from the other
• Lexical Studies – word-level alignments
• Tokenization Evaluation – test subword tokenizers on diverse domains
• Linguistic Research – syntax, morphology, semantics

📜 License

Apache 2.0

🤝 Citation

@dataset{tajpers_parallel_corpus_full_2026,
    title = {TajPersParallelCorpusFull} – A Unified Tajik–Persian Parallel Corpus,
    author = {TajikNLPWorld},
    year = {2026},
    publisher = {Hugging Face},
    url = {https://huggingface.co/datasets/TajikNLPWorld/TajPersParallelCorpusFull}
}

📬 Contact

Open an issue on the Hugging Face repository or contact the TajikNLPWorld team.

---

Generated by merging TajPersParallelCorpus and TajPersParallelLexicalCorpus with deduplication.