Neobe/dhivehi-byt5-latin2thaana-v1

Dhivehi ByT5 Latin to Thaana (v1)

This model is a specialized Latin-to-Thaana transliteration model optimized for Maldives news media, but also capable of handling both formal journalistic text and casual social media typing.

It functions as a "Hybrid" model: it respects the casual spacing of chat messages while correctly applying formal grammatical rules (such as compounding verbs) when it detects a news context.

🧠 Training Strategy

1. Base Model: google/byt5-small
2. General Fine-tuning: Tuned on the alakxender/dhivehi-transliteration-pairs dataset (~150k pairs) to learn general phonetics and spelling.
3. Domain Adaptation: Further fine-tuned on a high-quality dataset of 10k News headlines to recognize formal entities and apply correct grammatical spacing for official terms.

📊 Performance Samples

Category	Latin Input	Model Output
Formal Grammar	Raeesul jumhooriyya... thasdheegu kuravvaifi	ރައީސުލް ޖުމްހޫރިއްޔާ... ތަސްދީގުކުރައްވައިފި
Casual / Chat	Aharen miadhu varah ban'du hai	އަހަރެން މިއަދު ވަރަށް ބަނޑުހައި
Official Titles	Minister of Foreign Affairs Moosa Zameer	މިނިސްޓަރު އޮފް ފޮރިން އެފެއާސް މޫސަ ޒަމީރު
News Phrasing	Police service in vanee ekan kuhveri kohfa	ޕޮލިސް ސާވިސްއިން ވަނީ އެކަން ކުށްވެރިކޮށްފައި

💻 Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Neobe/dhivehi-byt5-latin2thaana-v1")
model = AutoModelForSeq2SeqLM.from_pretrained("Neobe/dhivehi-byt5-latin2thaana-v1")

text = "Raeesul jumhooriyya miadhu ganoonu thasdheegu kuravvaifi"
inputs = tokenizer(text, return_tensors="pt")

outputs = model.generate(**inputs, max_length=256)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# Output: ރައީސުލް ޖުމްހޫރިއްޔާ މިއަދު ގާނޫނު ތަސްދީގުކުރައްވައިފި