glassbox/shlyokavitsa-pairs

Shlyokavitsa → Cyrillic restoration pairs

210,236 (Latin, Cyrillic) phrase pairs for restoring shlyokavitsa (Bulgarian typed on a Latin keyboard) back into Cyrillic. Built from Bulgarian Wikipedia, so it can be shared under the same licence as its source.

{"latin": "sreshta se na dalbochina okolo", "cyrillic": "среща се на дълбочина около", "n_words": 5, "page_id": 1041}

Filed under translation because that is the closest category the Hub offers, but the task is script restoration within one language, not translation between two: input and target are the same Bulgarian sentence, one typed on a Latin keyboard and one in Cyrillic.

As far as we can tell this is the first shlyokavitsa dataset on the Hub; a Hub search for the term returns nothing else.

Loading

from datasets import load_dataset

pairs = load_dataset("glassbox/shlyokavitsa-pairs")
pairs["test"].filter(lambda r: r["n_words"] == 20)   # the length the models are scored at

n_words is the phrase length (3, 5, 10, 15, 20 or 25) and is the axis every number below is broken down along. page_id is the source Wikipedia article, which is also what the splits are assigned on.

Why the task is not a lookup table

The romanization is lossy in exactly the ways real shlyokavitsa is, so restoring it is a decision problem:

collision	example	share of pairs
a spells both а and ъ	dalbochina → дълбочина	45%
y spells both й and ь	neyniyat → нейният	27%
digraphs collide with letter pairs (sh, sht, ch, zh, yu, ya, c)	sreshta → среща	80%

Some phrases are irrecoverable in principle rather than merely hard. sasht is both САЩ (the abbreviation for the USA) and същ ("same"); faram is both фаръм (a transliterated place name) and фарам. A model that gets these wrong is not malfunctioning: the information is gone from the input. Any accuracy number on this dataset has that ceiling built into it.

How it was built

Source: HuggingFaceFW/finewiki, config bgwiki (Bulgarian Wikipedia, 284,820 articles).

1. Drop headings, list markers and short lines; split the remaining prose into sentences.
2. Lowercase and keep Cyrillic-only words, so the target alphabet is exactly the 30 letters plus space.
3. Sample word windows of 3, 5, 10, 15, 20 and 25 words within a single sentence, so no phrase crosses a sentence boundary.
4. Romanize each window with the fixed table below; drop exact duplicates (14,313 dropped).

а a  б b  в v  г g  д d  е e  ж zh  з z  и i  й y  к k  л l  м m  н n  о o
п p  р r  с s  т t  у u  ф f  х h  ц c  ч ch  ш sh  щ sht  ъ a  ь y  ю yu  я ya

Generator: datatools/make_shlyokavitsa_pairs.py, seed 0, deterministic. The same romanization table is what the restorer model was trained against, imported rather than copied so the two cannot drift.

Splits

Assigned per article, not per pair, so no sentence from a training article can reappear in validation or test:

split	pairs
train	200,097
validation	5,125
test	5,014

By phrase length: 57,517 (3 words), 58,715 (5), 43,243 (10), 27,140 (15), 15,422 (20), 8,199 (25). The generator stops reading as soon as every split hits its target size, and that happened after 5,763 articles (roughly 6 row groups) of the 285k available, so the recipe leaves the source almost entirely untouched if more is wanted.

A baseline, measured

A 4.73M-parameter character-level GPT trained on pairs generated the same way but from a different corpus (books, news, forum prose), scored on this test split:

phrase length	3	5	10	15	20	25
exact match, plain decoding	0.955	0.915	0.810	0.765	0.670	0.215
exact match, constrained decoding	0.965	0.920	0.820	0.790	0.720	0.370
character error rate, constrained	0.0021	0.0026	0.0036	0.0026	0.0031	0.0131

200 phrases per length, greedy decoding. Constrained decoding masks the decoder to the readings the input actually permits: zero parameters, and it is the row to beat. The same model scores 0.92 exact match at 20 words on in-corpus pairs, so this dataset is substantially harder: Wikipedia is dense with proper nouns, place names and abbreviations, which is where the irrecoverable collisions above concentrate.

For a second reference point, two general Bulgarian LLMs 550× and 900× larger (INSAIT's BgGPT-Gemma-2-2.6B-IT-v1.0 and BgGPT-Gemma-3-4B-IT, prompted 5-shot, greedy) score 0.095 / 0.045 / 0.015 / 0.015 / 0.005 / 0.000 and 0.090 / 0.060 / 0.025 / 0.015 / 0.005 / 0.000 across the same six lengths, same n = 200. The task is not something Bulgarian language modelling gives you for free at scale, which is what makes it worth a benchmark.

Limits

• Synthetic romanization, not observed typing. Real users are inconsistent: they mix 4 for ч, 6 for ш, q for я, drop diacritics unpredictably, and make typos. This dataset applies one fixed table. A model trained only on it will meet input it has never seen.
• Wikipedia register. Encyclopedic prose, heavy on named entities, light on the conversational text where shlyokavitsa is actually used.
• Lowercase only, no punctuation. The target alphabet is 30 letters plus space, matching the restorer's vocabulary. Casing and punctuation restoration are separate problems.
• Windows can start mid-clause, because phrases are sampled inside sentences rather than aligned to clause boundaries.

Licence and attribution

CC BY-SA 4.0, inherited from Bulgarian Wikipedia via FineWiki (which is CC BY-SA 4.0 and GFDL). Share-alike applies: derivatives carry the same licence. Each row keeps its source page_id; the article is https://bg.wikipedia.org/?curid={page_id}, and its authors are listed in that page's history.

Citation

@misc{shlyokavitsa_pairs,
  title     = {Shlyokavitsa to Cyrillic restoration pairs},
  author    = {Glassbox},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/glassbox/shlyokavitsa-pairs},
  note      = {Derived from HuggingFaceFW/finewiki (bgwiki), CC BY-SA 4.0}
}