Bashkir-Russian Parallel Corpus is a large-scale sentence-aligned parallel corpus for the Bashkir–Russian language pair, assembled from authentic human-created translations. It contains 3,040,085 unique parallel sentence pairs, where each Bashkir sentence is aligned with its Russian counterpart.

The corpus combines data from three open parallel corpora: TIL-MT Corpus (Turkic Interlingua), NLLB (OPUS), and Wikimedia (OPUS). All pairs were filtered by semantic similarity to remove misaligned and noisy entries, keeping only high-confidence translation pairs. Full duplicates were removed during compilation.

It is designed to support NLP research and applications for Bashkir, a low-resource Turkic language.

Curated by: BashkirNLPWorld
Language(s) (NLP): Bashkir (ba), Russian (ru), Cyrillic script
License: other – see Licensing below.

What BashkirNLPWorld contributed

The BashkirNLPWorld team built this dataset end-to-end on top of open sources:

Aggregation of three independent parallel corpora into a single unified dataset.
Field normalization to a minimal two-column schema (rus, bak).
Semantic filtering using multilingual sentence embeddings (cosine similarity).
Deduplication by the key (rus, bak) — full duplicates removed, 0 duplicates in the final release.
Noise filtering — removal of identical strings and pairs with anomalous length ratios.
Machine-readable packaging as UTF-8 JSONL, ready for datasets.load_dataset.
Documentation of sources, filtering criteria, statistics, and dataset limitations.

Dataset Sources

Repository: https://huggingface.co/datasets/BashkirNLPWorld/bashkir-russian-parallel
Original sources:
- TIL-MT Corpus — Turkic Interlingua parallel corpus (ba–ru)
- NLLB (OPUS) — NLLB ba–ru parallel corpus from opus.nlpl.eu
- Wikimedia (OPUS) — Wikimedia ba–ru parallel corpus from opus.nlpl.eu
Related datasets:

Licensing

Original sentence pairs belong to their respective source corpora (TIL-MT, NLLB/OPUS, Wikimedia/OPUS) and their contributors. BashkirNLPWorld does not claim ownership of this content.
The structured compilation, metadata, unified schema, deduplication, filtering, and documentation created by BashkirNLPWorld are released under CC BY-SA 4.0.
Users are responsible for complying with applicable law and the licenses of the original sources when reusing the data.
Rights holders can request removal of specific entries by contacting the maintainers — we respond within 14 business days and remove disputed content in the next release.

Uses

Direct Use

Training and fine-tuning machine translation systems (Bashkir ↔ Russian)
Cross-lingual representation learning and multilingual model training
Linguistic research on Bashkir–Russian translation patterns
Evaluation of translation quality and semantic alignment
Data augmentation for low-resource Turkic NLP
Instruction tuning of generative LLMs for Bashkir

Out-of-Scope Use

Re-licensing the original sentence pairs under a different license
Treating the dataset as a gold-standard translation reference without verification against the original sources
Any use that violates the rights of the original content creators

Dataset Structure

Data Fields

Field	Type	Description	Missing
`rus`	string	Russian sentence	0%
`bak`	string	Bashkir sentence	0%

Data Splits

The dataset contains a single split (train) with all 3,040,085 examples. Users are encouraged to create their own train/validation/test splits.

Example Record

{
  "rus": "Будто надсмехаясь над ним, повстанцы оглядываются назад.",
  "bak": "Көлгән һымаҡ, артҡа әйләнеп ҡарайҙар."
}

Dataset Statistics

Overview

Metric	Records	%
Total records	3,040,085	100.00%
Unique `(rus, bak)` pairs	3,040,085	100.00%
Full duplicates	0	0.00%

Semantic Similarity (`static_sim`)

Statistic	Value
Count	3,040,085
Min	0.2500
Max	1.0000
Mean	0.6270
Median	0.6393
Std Dev	0.1923
p1	0.2614
p5	0.3031
p25	0.4708
p50	0.6393
p75	0.7887
p95	0.9130
p99	0.9708

Russian Sentence Length (`length_rus`, characters)

Statistic	Value
Count	3,040,085
Min	1
Max	511
Mean	84.65
Median	66
Std Dev	65.25
p1	12
p5	19
p25	39
p50	66
p75	109
p95	213
p99	329

Bashkir Sentence Length (`length_bak`, characters)

Statistic	Value
Count	3,040,085
Min	1
Max	511
Mean	84.79
Median	68
Std Dev	62.56
p1	12
p5	21
p25	43
p50	68
p75	107
p95	208
p99	320

Russian Length Distribution (bins)

Range (chars)	Records	%
0–20	153,605	5.05%
20–40	612,774	20.16%
40–60	580,963	19.11%
60–80	459,889	15.13%
80–100	342,592	11.27%
100–150	489,456	16.10%
150–200	215,316	7.08%
200–300	140,162	4.61%
300–512	45,328	1.49%

Pairs by Semantic Similarity Threshold

Threshold	Remaining Pairs	% of total
`sim >= 0.25`	3,040,085	100.00%
`sim >= 0.5`	2,162,120	71.12%
`sim >= 0.6`	1,713,096	56.35%
`sim >= 0.7`	1,219,305	40.11%
`sim >= 0.8`	698,411	22.97%
`sim >= 0.9`	198,125	6.52%

Dataset Creation

Curation Rationale

Bashkir is a low-resource Turkic language with limited parallel data. The goal was to consolidate the most reliable openly available Bashkir–Russian parallel corpora into a single clean, deduplicated, and filtered resource suitable for machine translation and linguistic research.

Data Collection and Processing

Collected parallel data from three sources:
- TIL-MT Corpus (Turkic Interlingua, ba–ru)
- NLLB (OPUS) (ba–ru)
- Wikimedia (OPUS) (ba–ru)
Normalized the schema to two fields: rus, bak.
Filtered by semantic similarity using multilingual sentence embeddings.
Removed pairs where rus == bak (non-translations, duplicated strings).
Removed pairs shorter than 5 characters.
Removed pairs with length ratio greater than 3 (anomalous alignment).
Deduplicated by (rus, bak).
Saved as UTF-8 JSONL.

Tools: Python 3, Hugging Face datasets, sentence-transformers.

Source Data Producers

The original sentence pairs were produced by the authors, compilers, and contributors of the TIL-MT Corpus, NLLB, and Wikimedia projects. BashkirNLPWorld is responsible for aggregation, filtering, deduplication, unified schema, metadata, and packaging.

Annotations

No manual annotations were added. All sentence pairs come from the original parallel corpora.

Personal and Sensitive Information

The dataset contains publicly available sentence pairs and does not intentionally include personally identifiable information.

Bias, Risks, and Limitations

Source imbalance: the three sources contribute different volumes and domains; the corpus is not domain-balanced.
Domain skew: Wikimedia and NLLB data are skewed toward encyclopedic, religious, and technical texts.
Length skew: 44.3% of Russian sentences are shorter than 60 characters; long sentences (>300 chars) account for only 1.49%.
Semantic similarity skew: 28.9% of pairs have static_sim < 0.5, which may indicate loose or noisy translations.
Residual noise: despite semantic filtering, a small fraction of pairs may be imperfect translations (e.g., non-translated proper nouns such as "Любомир!" — "Любомир!").
No document context: sentences are aligned out of their original document context.
Source snapshot: reflects the state of the upstream corpora at collection time.

Recommendations

Account for semantic similarity distribution when training models — consider a higher threshold (static_sim >= 0.6 or >= 0.7) for higher precision.
Balance domains or sources if training a domain-specific MT system.
Filter out pairs with identical Russian and Bashkir strings before training.
For commercial use, verify the license status of specific source components.
Report issues via the Hugging Face repository.

Citation

BibTeX:

@dataset{bashkirnlpworld2026bashkirrussianparallel,
  author       = {BashkirNLPWorld},
  title        = {Bashkir-Russian Parallel Corpus},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/BashkirNLPWorld/bashkir-russian-parallel}
}

APA: BashkirNLPWorld. (2026). Bashkir-Russian Parallel Corpus [Data set]. Hugging Face. https://huggingface.co/datasets/BashkirNLPWorld/bashkir-russian-parallel

Glossary

Parallel corpus – a collection of texts aligned across two or more languages at the sentence level.
Sentence alignment – the task of matching sentences across languages that are translations of each other.
Semantic similarity (static_sim) – cosine similarity between multilingual sentence embeddings, ranging from 0 to 1.
OPUS – an open collection of parallel corpora (opus.nlpl.eu).
NLLB – No Language Left Behind, a multilingual translation project by Meta AI.
TIL-MT – Turkic Interlingua Machine Translation, a project focused on Turkic languages.
Low-resource language – a language with limited digital resources and tools.

Dataset Card Authors

BashkirNLPWorld

Dataset Card Contact

For questions, takedown requests, or collaboration, please open an issue: https://huggingface.co/datasets/BashkirNLPWorld/bashkir-russian-parallel/discussions

Downloads last month: 14

Total file size:

1.22 GB

You need to agree to share your contact information to access this dataset

Dataset Card for Bashkir-Russian Parallel Corpus

Dataset Details

Dataset Description