README.md

---
language:
- ky
- ru
license: cc-by-nc-nd-4.0
task_categories:
- translation
pretty_name: "KyRuBench-20K: Kyrgyz-Russian MT Benchmark"
size_categories:
- 10K<n<100K
tags:
- kyrgyz
- russian
- low-resource
- benchmark
- machine-translation
- evaluation
dataset_info:
  features:
  - name: source
    dtype: string
  - name: target
    dtype: string
  splits:
  - name: test
    num_examples: 20000
extra_gated_heading: "Request access to KyRuBench-20K"
extra_gated_description: "Access is granted automatically upon agreeing to the terms."
extra_gated_prompt: >-
  By requesting access to KyRuBench-20K, you agree to the following terms:

  1. You will use this dataset for **non-commercial research and educational purposes only**.

  2. You will **cite the dataset** in any publications, reports, or systems that use it.

  3. You will **not redistribute** the dataset or any derivative of it without prior written permission from the authors.

  4. You will **not create derivative datasets** based on this benchmark without permission.

  5. You will use this dataset **responsibly** and in accordance with applicable laws.

  6. You accept that the authors provide this dataset **as-is** and bear no liability for downstream use.

  7. You agree that your access information (name, email, affiliation) may be stored by the dataset maintainers for record-keeping purposes.
extra_gated_fields:
  Full Name: text
  Email: text
  Affiliation / Institution: text
  Country: country
  Intended use:
    type: select
    options:
      - Academic research
      - MT system development
      - LLM evaluation
      - Education
      - label: Other
        value: other
  Brief description of your use case: text
  I agree to cite this dataset in any resulting publications: checkbox
  I agree to the terms and conditions above: checkbox
extra_gated_button_content: "Agree and request access"
---

# KyRuBench-20K: A Domain-Balanced Benchmark for Kyrgyz→Russian Machine Translation

## Overview

KyRuBench-20K is a curated evaluation benchmark for Kyrgyz (кыргызча) to Russian (русский) machine translation. It contains **20,000 parallel sentence pairs** drawn from three distinct domains with **equal representation**, enabling fair cross-domain evaluation of translation systems.

This benchmark addresses the lack of standardized, balanced evaluation resources for Kyrgyz — a low-resource Turkic language with ~5 million speakers.

## Dataset Summary

| Property | Value |
|----------|-------|
| Language pair | Kyrgyz (ky) → Russian (ru) |
| Total examples | 20,000 |
| Domains | 3 (equally balanced) |
| Script | Cyrillic (both languages) |
| Split | Test only (evaluation benchmark) |

### Domain Distribution

| Domain | Examples | Description |
|--------|----------|-------------|
| **General Language** | 6,667 | Parallel text from news, government, and general web sources |
| **Mighty Kyrgyz (Yudahin)** | 6,666 | Yudahin dictionary-based parallel corpus |
| **Literature** | 6,667 | Literary text (novels, short stories, poetry) |

## Quality Filtering

The benchmark was constructed through a rigorous multi-stage filtering pipeline:

1. **Deduplication** — removed duplicate source and target sentences
2. **Length filtering** — retained sentences with 20–300 characters on both sides
3. **Length ratio** — source/target character ratio between 0.5 and 2.0
4. **Script validation** — Cyrillic ratio ≥ 0.6 on both sides
5. **Domain balancing** — equal random sampling from each domain (seed=42)

After filtering, 20,000 sentences were sampled with equal domain representation.

## Baseline Results

We evaluate 7 systems across three categories: dedicated MT, frontier LLMs, and web translators. All LLM-based systems use identical translation prompts for fair comparison.

| Rank | System | Type | BLEU | chrF++ | General | Yudahin | Literature |
|------|--------|------|------|--------|---------|---------|------------|
| 1 | AirUn Translator + LLM | Dedicated MT | **53.32** | **71.81** | 45.92 | 64.61 | 52.27 |
| 2 | AirUn Translator | Dedicated MT | 52.37 | 71.29 | 43.18 | 64.67 | 52.95 |
| 3 | Claude Opus 4.6 | Frontier LLM | 45.49 | 67.83 | 48.99 | 49.64 | 36.66 |
| 4 | GPT-5.4 | Frontier LLM | 43.53 | 66.59 | 47.58 | 48.60 | 33.68 |
| 5 | Aitil Translate | Web Translator | 42.15 | 64.35 | 46.91 | 47.27 | 31.33 |
| 6 | Claude Sonnet 4.6 | Frontier LLM | 38.85 | 63.75 | 41.56 | 44.97 | 30.08 |
| 7 | Grok 4.20 | Frontier LLM | 37.39 | 62.53 | 40.63 | 43.39 | 27.92 |

### Key Findings

- **Dedicated MT systems significantly outperform frontier LLMs** on Kyrgyz→Russian, with a gap of 8–16 BLEU points
- **Literature domain exposes the largest quality gap** between dedicated MT (BLEU ~52) and frontier LLMs (BLEU ~28–37)
- **General Language domain** is the most competitive, with some LLMs approaching dedicated MT performance
- **Claude Opus 4.6** is the strongest LLM for Kyrgyz, followed by GPT-5.4
- Kyrgyz remains a challenging low-resource language even for state-of-the-art LLMs

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("BDigit/KyRuBench-20K")

# Access the test split
for example in dataset["test"]:
    print(example["source"])   # Kyrgyz sentence
    print(example["target"])   # Russian reference translation
```

## Evaluation

We recommend evaluating with [sacreBLEU](https://github.com/mjpost/sacrebleu) for reproducibility:

```python
import sacrebleu

# After generating hypotheses
bleu = sacrebleu.corpus_bleu(hypotheses, [references])
chrf = sacrebleu.corpus_chrf(hypotheses, [references], word_order=2)  # chrF++

print(f"BLEU: {bleu.score:.2f}")
print(f"chrF++: {chrf.score:.2f}")
```

**Important:** Always report per-domain scores alongside overall metrics. The three domains test fundamentally different translation challenges.

## Data Fields

- **source** (`str`): Sentence in Kyrgyz (Cyrillic script)
- **target** (`str`): Reference translation in Russian (Cyrillic script)

## Intended Use

- Benchmarking machine translation systems for Kyrgyz→Russian
- Evaluating multilingual LLMs on low-resource Turkic language translation
- Comparing dedicated MT systems against general-purpose language models
- Studying domain effects on translation quality for low-resource languages

## Limitations

- **Single reference**: Each source has one reference translation. Multi-reference evaluation would give more robust scores.
- **Direction**: Currently Kyrgyz→Russian only. Russian→Kyrgyz evaluation requires separate references.
- **Literature domain**: Literary text availability is limited, constraining future expansion of this domain.
- **Reference quality**: References may contain occasional noise from the original parallel corpus.

## Citation

```bibtex
@misc{kyrubench2026,
  title={KyRuBench-20K: A Domain-Balanced Benchmark for Kyrgyz-Russian Machine Translation},
  author={Alibekov, Nurtilek and Kumarbai uulu, Bektemir and Uvalieva, Zarina and Tashbaltaev, Tynchtykbek and Metinov, Adilet},
  year={2026},
  organization={Bdigital LLC},
  howpublished={\\url{https://huggingface.co/datasets/BDigit/KyRuBench-20K}},
}
```

## License

CC-BY-NC-ND-4.0 (Attribution, Non-Commercial, No Derivatives)

For commercial licensing inquiries, contact an@bdigital.kg.