Datasets:

blinoff
/

UzLPC

license: cc-by-sa-4.0
task_categories:
  - translation
language:
  - uz
  - en
  - ru
tags:
  - parallel corpus
  - tranlate
pretty_name: UzLPC
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.csv
    sep: "\t"
hosted_by_space: blinoff/UzLPC

🇺🇿 Uzbek Large-Scale Parallel Corpus 2M (UzLPC)

The largest open parallel corpus for the Uzbek language Combining Tatoeba and Wikipedia data with high-quality translations from Qwen3.5-122B-A10B.

📖 Description

The Uzbek Large-Scale Parallel Corpus 2M (UzLPC) is a comprehensive collection of over 2.3 million parallel sentences, designed to advance NLP research for the Uzbek language. It includes:

Source texts in English and Russian, with translations into Uzbek.
Reference alignments (for Wikipedia portion) and confidence scores to support machine translation, cross-lingual transfer, and linguistic analysis.
Agglutination indices to facilitate morphological and typological studies of Uzbek.

The corpus is built from two high-quality sources:

Tatoeba – a community-driven database of example sentences and translations (89.6% of the data).
Uzbek Wikipedia – aligned with English articles using LaBSE embeddings, providing high-confidence parallel pairs (10.4% of the data).

All Uzbek translations were generated using the Qwen3.5-122B-A10B model (MoE, 122B parameters) running on an Ascend 910B NPU cluster, ensuring fluency and consistency.

📊 Dataset Statistics

Metric	Value
Total sentence pairs	2,334,217
from Wikipedia	242,730 (10.40%)
from Tatoeba	2,091,487 (89.60%)
Source language distribution	English: 1,488,583 (63.77%) Russian: 845,634 (36.23%)

Word‑level statistics (Uzbek translations)

Metric	Value
Total words	17,965,600
Unique words (vocabulary)	819,160
Type–Token Ratio (TTR)	0.0456
Average source word length	4.78 characters
Average translation word length	6.69 characters

Sentence‑level statistics

Language	Mean word count	Median word count
Source (English/Russian)	8.93	7.00
Translation (Uzbek)	7.70	6.00

Translate metrics

On Wikipedia part: BLEU=16.19, chrF=58.48

🗂️ Data Structure

The corpus is provided as a single CSV file with the following columns:

Column	Description
`sentence_id`	Unique identifier (original sentence ID from Tatoeba courpus or Wikipedia sentence ID)
`source_lang`	Source language code (`eng` or `rus`)
`source_texts`	Original sentence in English or Russian
`translation`	Uzbek translation generated by Qwen3.5-122B-A10B
`references`	Reference Uzbek sentence (available only for Wikipedia part)
`confidence_score`	Semantic similarity score (LaBSE) between source and reference (Wiki only)
`agglutination_index`	Morphological agglutination index calculated for the reference sentence (Wiki only)

Note: For Tatoeba entries, references, confidence_score, and agglutination_index are left empty.

🛠️ Methodology

Data Collection
- Tatoeba: Downloaded the full sentence dump, filtered for English and Russian source sentences.
- Wikipedia: Extracted articles from the Uzbek Wikipedia dump and aligned them with their English counterparts.
Translation
- All source sentences (English and Russian) were translated into Uzbek using the Qwen3.5-122B-A10B model.
- Translation was performed on a high‑performance server with 8× Ascend 910B NPUs (512 GB VRAM) using SGLang, achieving an average throughput of >160 tokens/second.
Alignment (Wikipedia)
- Sentence‑level alignment between Uzbek and English Wikipedia articles was performed using LaBSE embeddings and cosine similarity.
- Only pairs with a confidence score ≥ 0.7 were retained, yielding 242,730 high‑quality parallel sentences.
Morphological Analysis
- The agglutination index was computed for each Uzbek reference sentence to quantify the degree of agglutination, a key typological feature of the language.

🙏 Acknowledgments

We extend our sincere gratitude to Huawei Technologies for providing the computing resources necessary for this project. The translation of over 2.3 million sentences was performed on an Atlas 800I A2 server equipped with 8× Ascend 910B NPUs (512 GB VRAM) and 4× Kunpeng 920 CPUs. This powerful infrastructure enabled us to achieve an inference throughput exceeding 160 tokens/second, making this large-scale corpus creation possible.

We also thank the open-source community for maintaining the Tatoeba project and Wikipedia, which served as the foundation of this dataset.

📜 License

This dataset is released under the Creative Commons Attribution‑ShareAlike 4.0 International (CC BY‑SA 4.0) license, consistent with the licenses of its source data (Tatoeba and Wikipedia).
You are free to use, share, and adapt the dataset, provided you give appropriate credit and distribute any derivative works under the same terms.