Access is gated with manual approval. Users must request access on Hugging Face, and the repository owner must approve the request before files can be downloaded.

No test set is included in this release.

Data Summary

Data type	Total sentences	Details
Bilingual training data	140,000	20,000 sentence pairs for each of 7 language pairs.
Bilingual validation data	14,000	2,000 sentence pairs for each of 7 language pairs.
Monolingual training data	683,416	Source-language monolingual data for 7 languages, supplemented where additional monolingual source files were available.

Bilingual Configurations

Each bilingual config contains a train split and a validation split.

Source language	Config	Train rows	Validation rows	Train file	Validation file
Indonesian	`id-zh`	`20,000`	`2,000`	`id-zh-train.csv`	`id-zh-valid.csv`
Thai	`th-zh`	`20,000`	`2,000`	`th-zh-train.csv`	`th-zh-valid.csv`
Khmer	`km-zh`	`20,000`	`2,000`	`km-zh-train.csv`	`km-zh-valid.csv`
Malay	`ms-zh`	`20,000`	`2,000`	`ms-zh-train.csv`	`ms-zh-valid.csv`
Vietnamese	`vi-zh`	`20,000`	`2,000`	`vi-zh-train.csv`	`vi-zh-valid.csv`
Burmese	`my-zh`	`20,000`	`2,000`	`my-zh-train.csv`	`my-zh-valid.csv`
Lao	`lo-zh`	`20,000`	`2,000`	`lo-zh-train.csv`	`lo-zh-valid.csv`

Bilingual Schema

source_content
target_content
target_lang_char_count
domain

Monolingual Configurations

Each monolingual config contains a train split.

Language	Config	Train rows	Train file
Indonesian	`id-mono`	`98,000`	`id-mono-train.csv`
Thai	`th-mono`	`98,000`	`th-mono-train.csv`
Khmer	`km-mono`	`97,997`	`km-mono-train.csv`
Malay	`ms-mono`	`96,178`	`ms-mono-train.csv`
Vietnamese	`vi-mono`	`98,000`	`vi-mono-train.csv`
Burmese	`my-mono`	`97,251`	`my-mono-train.csv`
Lao	`lo-mono`	`97,990`	`lo-mono-train.csv`

Monolingual Schema

text
domain

Split Construction

The previous high and low files were merged per language pair. A fixed random seed was used to sample 20,000 bilingual training rows and 2,000 bilingual validation rows per language pair. Remaining source-language rows were converted into monolingual training data. Khmer and Lao were supplemented from additional source sentence spreadsheets; Indonesian, Thai, Malay, Burmese, and Vietnamese were supplemented from recrawled monolingual files. A separate monolingual test repository was created by sampling 2,000 rows per language from the monolingual pool and removing those rows from monolingual training files.

Downloads last month: 72

Total file size:

2.21 GB