ghananlpcommunity/pristine-twi-english-bidirectional-multilen

Twi ↔ English Bidirectional MT Dataset

Derived from ghananlpcommunity/pristine-twi-english.

Schema

Field	Type	Description
source	string	Source text (no prefix token)
target	string	Clean target translation
direction	string	twi-en or en-twi
style	string	plain or paragraphs — format control lives here
num_sentences	int	Number of sentences in the example

Format-control

The style column tells the model what output format to produce. At inference time, prepend a token yourself if your model uses them, or use the style column directly as a conditioning signal.

style value	Meaning
plain	Flat, continuous text
paragraphs	Paragraph-formatted output (blank lines between paragraphs)

Design

• Sentence lengths 1–{max_sents} are all represented, each with exactly {balance_target} examples per direction.
• Both translation directions (twi→en, en→twi) are balanced 50/50.
• ~{int(PARAGRAPH_FRACTION*100)}% of examples with ≥{MIN_SENTS_FOR_PARAGRAPHS} sentences use paragraph formatting.
• Paragraphs contain {MIN_GRAF_SIZE}–{MAX_GRAF_SIZE} sentences each.

Splits

Split	Examples
train	{len(train_data):,}
validation	{len(val_data):,}
test	{len(test_data):,}
"""