deshanksuman/swabhashambart50SinhalaTransliteration

Swa bhasha mBART-50 Sinhala Transliteration Model

This model transliterates Romanized Sinhala text to Sinhala script.

Model description

This is a fine-tuned version of deshanksuman/mbart_50_SinhalaTransliteration specialized for Sinhala transliteration. It converts romanized Sinhala (using Latin characters) to proper Sinhala script. Due to the training limitation only 2/3 of data was utilized for the training purposes.

Intended uses & limitations

This model is intended for transliterating Romanized Sinhala text to proper Sinhala script. It can be useful for:

• Text input conversion in applications
• Helping non-native speakers type in Sinhala
• Converting legacy text in romanized format to proper Sinhala

Acknowledgement

We acknowledge the support of the Supercomputing Wales project, which is part-funded by the European Regional Development Fund (ERDF) via Welsh Government.

How to use

from transformers import MBartForConditionalGeneration, MBartTokenizer

# Load model and tokenizer
model_name = "deshanksuman/swabhashambart50SinhalaTransliteration"
tokenizer = MBartTokenizer.from_pretrained(model_name)
model = MBartForConditionalGeneration.from_pretrained(model_name)

# Set language codes
tokenizer.src_lang = "en_XX"  # Using English as source language token
tokenizer.tgt_lang = "si_LK"  # Sinhala as target

# Prepare input
text = "mama oyata adare karanawa"
inputs = tokenizer(text, return_tensors="pt", max_length=128, padding="max_length", truncation=True)

# Generate output
outputs = model.generate(
    input_ids=inputs["input_ids"],
    attention_mask=inputs["attention_mask"],
    max_length=128,
    num_beams=5,
    early_stopping=True
)

# Decode output
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

Training data

The model was trained on the deshanksuman/Augmented_SinhalatoRomanizedSinhala_Dataset dataset, which contains pairs of Romanized Sinhala and corresponding Sinhala script text.

Training procedure

The model was trained with the following parameters:

• Learning rate: 5e-05
• Batch size: 32
• Number of epochs: 1
• Max sequence length: 128
• Optimizer: AdamW

Model Performance

The model achieves an accuracy of 72.00% on the test set.

Examples:

Example 1:

• Romanized: karandeniya em
• Expected: කරන්දෙණිය එම්
• Predicted: කරන්දෙණිය එම්
• Correct: True

Example 2:

• Romanized: yatawena minissu hoya ganna beriwa mihidan wenawa
• Expected: යටවෙන මිනිස්සු හොයා ගන්න බැරිව මිහිදන් වෙනවා
• Predicted: යටවෙන මිනිස්සු හොයා ගන්න බැරිව මිහිදන් වෙනවා
• Correct: True