sukhrobnurali/uzbek-embedding-pairs

Uzbek Embedding Pairs

Parallel and monolingual Uzbek sentence pairs for training sentence-embedding models. Combines OPUS-100, Uzbek Wikipedia, FLORES-200 dev, and 10% Cyrillic-Latin mix-script augmentation. Generated 2026-05-17.

Repository: sukhrobnurali/uzbek-embedding-pairs

Dataset Structure

Config	Split	Rows	Use
default	train	356,278	Training pool (parallel + monolingual + mixscript aug)
default	validation	1,994	In-loop eval during training (FLORES-200 dev)
wiki_retrieval_eval	test	5,000	Held-out Wikipedia title-paragraph pairs for retrieval eval
smoke_100	train	100	Stratified 100-row sample for pipeline integrity tests

Schema

Every row has the same five columns:

Column	Type	Description
anchor	string	First text in the pair (typically Uzbek for parallel sources, title for wiki)
positive	string	Paired text (typically English translation, or wiki paragraph, or transliterated Uzbek for mixscript)
source	string	Origin tag - see "Source Datasets" below
anchor_lang	string	uz_Latn | uz_Cyrl | en | mixed | empty (auto-detected)
positive_lang	string	Same vocabulary as anchor_lang

Source Datasets

Source	HF ID	License	Role
OPUS-100 en-uz	Helsinki-NLP/opus-100	unknown (flagged below)	Parallel sentences (subset, after dedup)
Parallel-sentences-OPUS	sentence-transformers/parallel-sentences-opus-100	unknown	Cleaner OPUS variant (preferred on dedup collisions)
Uzbek Wikipedia	yakhyo/uz-wiki	MIT	Title-first-80-words pairs
FLORES+ dev	openlanguagedata/flores_plus	CC-BY-SA-4.0 (gated)	Validation split only (uzn_Latn-eng_Latn; Cyrillic side transliterated)

Source distribution in default/train

Source	Rows	Share
wiki	199,208	55.9%
parallel_opus	121,442	34.1%
mixscript	35,628	10.0%

Processing Pipeline

1. Load each source with utils.dataset_io.load_source(name) to a unified 5-column schema.
2. Normalize Unicode to NFC; canonicalize apostrophes (o', g' markers).
3. Filter: length 5-512 chars per side (2000 for wiki paragraphs), word count >= 2, length ratio 0.4-2.5 for parallel sources, drop identity rows, Uzbek-side script-sanity check.
4. Exact dedup on (anchor.lower(), positive.lower()).
5. Near dedup via MinHash LSH (joint key anchor || positive, threshold 0.90, num_perm 64). Joint key chosen over anchor-only after the anchor-only variant collapsed OPUS-100 (median 6-word anchors) too aggressively.
6. Mix-script augmentation: sample eligible Uzbek anchors stratified by source, transliterate to the opposite script via utils.translit.to_cyrillic / to_latin, emit (original, transliterated) pairs with source="mixscript". Target: 10.0% of final pool.
7. Validation split assembled from FLORES-200 uzn_Latn-eng_Latn dev plus the Cyrillic counterpart (see below).
8. Wiki retrieval holdout: 5,000 random articles (seed=42) split off before training and exposed as the wiki_retrieval_eval config.

Validation Split

The validation split of the default config is FLORES-200 dev - held out from training.

Cyrillic validation rows are transliterated from the Latin dev split via utils.translit.to_cyrillic (the public FLORES-200 release did not include uzn_Cyrl-eng_Latn at build time). Rows have source="flores_dev_cyrl_translit" for transparency.

flores_devtest is never included in any config - it is reserved for the downstream model evaluation script (eval.py).

Intended Use

• Training Uzbek sentence-embedding / retrieval models with contrastive losses (e.g., MultipleNegativesRankingLoss).
• Compatible with sentence-transformers SentenceTransformerTrainer.
• Anchor-positive is treated symmetrically - no query: / passage: prefix required for the IBM Granite r2 base.

Limitations

• Cyrillic Е/Э one-way mapping (rule-based fallback). The rule-based transliterator maps Latin E to Cyrillic Е only, never Э. Native Uzbek words overwhelmingly use Е; rare loanwords using Э lose the distinction on round-trip. UzTransliterator==0.0.36 is the primary path and handles this correctly when installed.
• OPUS-100 en-uz license is unknown. Included under common research-redistribution conventions. Downstream users in restrictive jurisdictions should verify the source license themselves.
• Wiki title overlaps in the retrieval holdout. Two articles in the 5,000-row holdout share a title with a training article (Wikipedia redirects/disambig pages). Minor leakage; flagged for transparency.
• Wiki pairing is heuristic. Title and first-80-words after dropping disambig pages and title-in-first-sentence rows. Some pairs are weakly grounded.
• auto_detect_script is heuristic. Latin text containing q/x is tagged uz_Latn (rare in English roots); pure English with these letters may be misclassified.

Citations

Please cite the source datasets you transitively use:

@inproceedings{zhang-etal-2020-improving,
  title     = {Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation},
  author    = {Zhang, Biao and Williams, Philip and Titov, Ivan and Sennrich, Rico},
  booktitle = {Proceedings of ACL},
  year      = {2020}
}

@article{nllb2022,
  title   = {No Language Left Behind: Scaling Human-Centered Machine Translation},
  author  = {NLLB Team},
  journal = {arXiv preprint arXiv:2207.04672},
  year    = {2022}
}

@misc{uzwiki2024,
  title  = {Uzbek Wikipedia dump},
  author = {Yakhyo, A.},
  year   = {2024},
  url    = {https://huggingface.co/datasets/yakhyo/uz-wiki}
}

@misc{granite_embedding_r2,
  title  = {Granite Embedding 97m Multilingual r2},
  author = {IBM Granite Team},
  year   = {2026},
  url    = {https://huggingface.co/ibm-granite/granite-embedding-97m-multilingual-r2}
}

License

Apache-2.0 for the processing pipeline and curation choices in this dataset. Source-row licenses are preserved per the row's source tag - consult the table above for upstream terms.

Reproduction

git clone https://github.com/sukhrobnurali/uzbek-embedding-pairs
cd uzbek-embedding-pairs
pip install -r requirements.txt
python prepare_data.py --push --repo-id <your-fork>

See Also

• Model trained on this dataset (forthcoming): sukhrobnurali/uzbek-granite-97m-v1
• Live demo Space (forthcoming): sukhrobnurali/uzbek-embedding-demo
• Base embedding model: ibm-granite/granite-embedding-97m-multilingual-r2