---
language:
- dv
tags:
- transliteration
- dhivehi
- thaana
- byt5
- news
- nlp
- latin-to-thaana
license: apache-2.0
base_model: google/byt5-small
datasets:
- alakxender/dhivehi-transliteration-pairs
widget:
- text: "Raeesul jumhooriyya miadhu ganoonu thasdheegu kuravvaifi"
  example_title: "Formal News (Grammar)"
- text: "Aharen miadhu varah ban'du hai"
  example_title: "Casual / Chat"
- text: "Minister of Foreign Affairs Moosa Zameer"
  example_title: "Official Titles"
inference:
  parameters:
    max_length: 256
---

# 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](https://huggingface.co/google/byt5-small)
2.  **General Fine-tuning:** Tuned on the [alakxender/dhivehi-transliteration-pairs](https://huggingface.co/datasets/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

```python
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