corpora (Mandarin - Taiwanese Hokkien)
Mandarin (zh, Traditional) and Taiwanese Hokkien (nan) text corpora built from the local ezai_corpus.
Configs
| config | folder | splits | rows |
|---|---|---|---|
parallel |
test/parallel/ |
train / eval / test | 412,884 / 1,928 / 3,198 |
mono_nan |
test/mono_nan/ |
train / eval | 1,073 / 100 |
kautian_test |
kautian_test/ |
train | 16,819 |
from datasets import load_dataset
ds = load_dataset("Jumping0629/corpora", "parallel") # or "mono_nan", "kautian_test"
parallel and mono_nan are the zh-nan-corpus demo split. Their columns, sources, cleaning and split policy
are documented in test/README.md, with the build scripts in test/scripts/.
kautian_test
Example sentences and words from the kautian dictionary corpus (ezai_original/v1/kautian-cut.json), with
only four columns kept. This is the raw v1 text: no Gemini cleaning, no dedup, no split.
| column | source field | description |
|---|---|---|
id |
例句id |
example sentence id (string, unique) |
HAN |
漢字 |
Taiwanese Hokkien in Hanzi |
TL |
羅馬字 |
Tâi-lô romanization |
ZH |
華語 |
Mandarin translation |
- 16,819 rows, in the same order as the source file, all in
train. - No empty
HAN/TL/ZHvalues. 162 rows repeat another row's (HAN,TL,ZH) exactly; they are kept. - Dropped source fields:
audio,華語(原始),漢字斷詞,羅馬字斷詞. - kautian is also part of
parallel(sourcekautian, using the cleaned text). Do not evaluate onparalleleval/test after training onkautian_test: the same sentences may appear in both.
Licensing
The license/usage rights of the underlying sources have NOT been confirmed. license: other is a placeholder and does not grant any rights.
Until each source's terms are verified, keep the repository private and gated, and do not redistribute.
- Downloads last month
- 807