Datasets:
Zo corpus text: clean plain text in 24 Zo / Kuki-Chin languages
Plain text for reading (continued pre-training), one configuration per language, 312,635,883 words in all. Built 11 November 2026 from web crawls, public datasets, scripture, scanned school books and the owner's own collections. Junk, English, other scripts and duplicates were removed, and the language of each document was checked.
from datasets import load_dataset
ds = load_dataset("sensix-zo/zo-corpus-text", "paite_pck") # one row per line of text
Each language folder has text.txt.gz (the text), wordlist.tsv (word frequencies) and SOURCES.md (where
the text came from and what was dropped). manifest.json has the counts for everything.
What is in it
| Language | Config | Words | Language label | Share that is Bible | Machine-made text set aside (words, not included) |
|---|---|---|---|---|---|
| Mizo (Lushai) | mizo_lus |
145,121,679 | verified | 0% | 359,297 |
| Tedim (Zolai) | tedim_ctd |
56,428,822 | verified | 3% | 46,144,251 |
| Hakha Chin (Lai) | hakha_cnh |
28,904,592 | verified | 0% | 31 |
| Paite | paite_pck |
27,003,091 | verified | 6% | 31,119,479 |
| Thadou-Kuki | thadou_tcz |
14,483,907 | verified | 9% | 171,789 |
| Hmar | hmar_hmr |
11,193,170 | verified | 7% | 4,153 |
| Falam Chin | falam_cfm |
8,520,359 | verified | 0% | 93 |
| Vaiphei | vaiphei_vap |
5,193,344 | verified | 17% | 10,512 |
| Zou | zou_zom |
4,089,479 | verified | 19% | 1,062 |
| Khumi Chin | khumi_cnk |
3,071,172 | tagger only | 0% | 17 |
| Gangte | gangte_gnb |
1,754,181 | verified | 1% | 12,148 |
| Siyin (Sizang) | siyin_csy |
1,541,269 | tagger only | 12% | 655 |
| Simte | simte_smt |
1,290,726 | verified | 60% | 14,634 |
| Bawm | bawm_bgr |
1,107,556 | tagger only | 0% | 37 |
| Matu Chin | matu_hlt |
911,911 | verified | 82% | 0 |
| Mara | mara_mrh |
515,048 | verified | 0% | 552 |
| Ngawn | ngawn_cnw |
279,070 | tagger only | 0% | 159 |
| Zyphe (Zophei) | zyphe_zyp |
278,463 | tagger only | 82% | 0 |
| Lamkang | lamkang_lmk |
212,795 | tagger only | 0% | 0 |
| Daai Chin | daai_dao |
204,534 | tagger only | 91% | 0 |
| Undetermined Zo language | undetermined |
152,638 | - | 0% | 0 |
| Zotung | zotung_czt |
145,547 | tagger only | 0% | 81 |
| Biate | biate_biu |
140,966 | tagger only | 0% | 645 |
| Kom | kom_kmm |
78,908 | tagger only | 0% | 1,022 |
| Aimol | aimol_aim |
12,656 | tagger only | 0% | 0 |
The word counts are from before that removal; Paite is now about 26.0 million words.
"verified" = the language label was checked against reliable text. "tagger only" = the label comes from a public crawl's automatic tagger and could not be checked: expect some wrong-language or junk pages.
Known problems
- Known machine-made sentences were taken out on 19 November 2026. Any whole line that is a known
machine-made sentence (a set of machine-translated short sentences, for example "Chibai zingkal" for "Good
morning", which is wrong; and the Zo side of every pair labelled
machine-translatedorllm-generatedinsensix-zo/zo-pairs) was removed: 109,756 lines from Paite (1,061,460 words) and 423 lines from all other languages together. Only exact whole-line matches are caught. A machine-made sentence inside a longer line, or one that is in no list, is still there. - Much of the small-language text is Bible (Simte 60%, Matu 82%, Daai 91%, Zyphe 82%), so models learn a church register.
- Text from scanned books (OCR) is good, not perfect.
- The Paite text includes vulgar and body vocabulary on purpose (about 1,400 lines), so a model knows those words.
Rights
A research collection. News, blogs and scripture remain their publishers' copyright; the Bharatavani school books are licensed for educational and research use only, non-commercial. Do not pass the files on.
- Downloads last month
- 686