chagatai-project/chagatai-sbd

Chagatai Sentence Boundary Detection

Canonical word-level Sentence Boundary Detection data for Chagatai. South Uzbek (uzs) and Uyghur (uig) are optional train-only auxiliary languages. Every configuration uses the same Chagatai train source split. Validation and test are physically shared files referenced by all five configurations.

Load with datasets

from datasets import load_dataset

dataset = load_dataset("chagatai-project/chagatai-sbd", "chagatai_only")
balanced = load_dataset(
    "chagatai-project/chagatai-sbd",
    "chagatai_uzs_uyghur_balanced",
)

train = balanced["train"]
validation = balanced["validation"]
test = balanced["test"]

The full multilingual configuration is large. It can be streamed without a complete local download:

full_train = load_dataset(
    "chagatai-project/chagatai-sbd",
    "chagatai_uzs_uyghur_full",
    split="train",
    streaming=True,
)

Configurations

Configuration	Train source sentences	Train sequences	Validation	Test
chagatai_only	Chagatai 3,035	3,045	145	295
chagatai_uyghur	Chagatai 3,035, Uyghur 3,035	6,090	145	295
chagatai_uzs	Chagatai 3,035, UZS 3,035	6,090	145	295
chagatai_uzs_uyghur_full	Chagatai 3,035, UZS 35,393, Uyghur 238,344	276,861	145	295
chagatai_uzs_uyghur_balanced	Chagatai 3,035, UZS 3,035, Uyghur 3,035	9,135	145	295

The pairwise and balanced configurations cap every auxiliary language at 3,035 source sentences, matching Chagatai train. This is sentence- and sequence-level balance, not token-level balance.

Warning: The chagatai_uzs_uyghur_full configuration is heavily skewed towards Uyghur (238,344 Uyghur source sentences vs. 3,035 Chagatai). For cross-lingual and balanced multilingual training, we strongly recommend using chagatai_uzs_uyghur_balanced.

Validation contains 433 Chagatai source sentences grouped into 145 sequential examples. Test contains 868 Chagatai source sentences grouped into 295 sequential examples.

Labels and columns

labels is aligned with tokens:

• 0: the word does not end a sentence;
• 1: the word is the final word of a real source sentence.

Each row also includes the cleaned text, language, augmentation method, source sentence IDs, fragment spans, boundary flags, summary counts, and where available from the source manuscript, Kazakh translation and transcript.

Construction rules

Chagatai is split in source order 70/10/20 before augmentation. Validation and test contain sequential Chagatai only. Train uses:

• sequential concatenation of 2-4 neighboring source sentences;
• random concatenation of 2-5 distinct, nonadjacent, reordered sources;
• partial merge of the end of one sentence and the start of another.

In partial examples, the end of the first fragment remains EOS and the end of the second fragment is not EOS because the original sentence continues.

The cleaning pass applies NFKC, removes punctuation, symbols, controls, and combining marks, and rejects residual Latin, Cyrillic, CJK, URL, and bibliography noise.

Validation and provenance

All five configurations passed source-split leakage, exact sequence overlap, label length, fragment reconstruction, auxiliary train-only, random ordering, and partial-boundary checks. Build manifests and statistics are stored under metadata/<configuration>/.

Cleaned source tables are available under sources/<configuration>/. They are not loaded as model splits, but retain source locators and raw text for audit and reconstruction.

The source corpora retain their original terms. No new license is asserted for the combined dataset; users should verify the terms of each source before redistribution or commercial use.

Sources and Citation

If you use this dataset, please cite the underlying sources:

• Chagatai (chg): Original historical texts sourced from KazCorpus / QazCorpora (National Corpus of the Kazakh Language), containing the manuscript Shajara-i Turk (Genealogy of the Turks) by Abu al-Ghazi Bahadur Khan (17th c.).
• South Uzbek (uzs): Auxiliary data sourced from the Hugging Face dataset tahrirchi/lutfiy (Tahrirchi, "Filling the Gap for Uzbek: Creating Translation Resources for Southern Uzbek", arXiv:2508.14586).