It is designed to handle the complexities of journalistic text, focusing on the correct preservation of context, English loan words, and named entities (e.g., Country names, Acronyms like UN, Sports teams).

🧠 Training Strategy

Base Model: google/byt5-small
General Fine-tuning: Tuned on the alakxender/dhivehi-transliteration-pairs dataset (~180k pairs) to learn general transliteration rules.
Domain Adaptation: Further fine-tuned on a high-quality dataset of 10k News headlines to capture journalistic style and entities.

📊 Performance Samples

Category	Thaana Input	Model Output
Global Politics	ސީރިއާގެ ކަންކަމުގެ ތެރެއަށް އިސްރާއީލް ނުވަނުމަށް	`Syria ge kankamuge thereah Israel nuvanumah`
Acronyms	ރާއްޖޭގެ ރެކޯޑުތައް އދ އިން ބަލަނީ	`Raajjeyge record thah UN in balanee`
English Loan Words	ސްކޭމާއި ދެކޮޅަށް ހިންގި ޚާއްސަ އޮޕަރޭޝަނެއް	`Scam aai dhekolhah hingi khaassa operation eh`
Sports	ވެސްޓް ހޭމް އަތުން ޗެލްސީ އަށް މޮޅެއް	`West Ham athun Chelsea ah molheh`
Dates	އޮކްޓޯބަރ 3 ގެ މުޒާހަރާ	`October 3 ge muzaaharaa`

💻 Usage

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Neobe/dhivehi-byt5-thaana2latin-v1")
model = AutoModelForSeq2SeqLM.from_pretrained("Neobe/dhivehi-byt5-thaana2latin-v1")

text = "ސީރިއާގެ ކަންކަމުގެ ތެރެއަށް އިސްރާއީލް ނުވަނުމަށް ޓްރަމްޕް ގޮވާލައްވައިފި"
inputs = tokenizer(text, return_tensors="pt")

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

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# Output: Syria ge kankamuge thereah Israel nuvanumah Trump govaalavvaifi