sin-transliterator 0.1.1

Sinhala transliteration via fine-tuned Transformer and LLM models, with optional code-mix support.

sin_transliterator

A lightweight Python package for Sinhala phonetic and adhoc transliteration — converting romanised Sinhala (Singlish) to Sinhala Unicode script. Backed by four fine-tuned models hosted on HuggingFace, with full support for Sinhala-English code-mixed input.

---

Table of Contents

• Overview
• Installation
• Quickstart
• Batch Transliteration
• Model Selection Guide
• VRAM & Hardware Requirements
• API Reference
• Error Handling
• Acknowledgements
• Contact

---

Overview

sin_transliterator exposes a single class — SinTransliterator — that resolves the correct model at runtime based on two parameters: the model architecture family and whether the input contains code-mixed text.

Four fine-tuned models underpin the package:

Model	Architecture	Code-Mix Support	HuggingFace
Small100 Base	Seq2Seq Transformer	No	savinugunarathna/Small100-Singlish-Sinhala-Merged
Small100 Code-Mix	Seq2Seq Transformer	Yes	savinugunarathna/small-100-Singlish-Sinhala-CodeMix
Gemma3 Base	Causal LLM	No	savinugunarathna/Gemma3-Singlish-Sinhala-Merged
Gemma3 Code-Mix	Causal LLM	Yes	savinugunarathna/Gemma3-Singlish-Sinhala-CodeMix

---

Installation

pip install sin_transliterator

---

Quickstart

Downloads model weights on first run and caches them locally. Works on any machine — GPU is used automatically if available, otherwise falls back to CPU.

from sin_transliterator import SinTransliterator

# Pure Sinhala input
t = SinTransliterator(model="transformer", contains_code_mix=False)
print(t.transliterate("mama yanawa"))

# Code-mixed input (Sinhala + English)
t_mix = SinTransliterator(model="transformer", contains_code_mix=True)
print(t_mix.transliterate("mama office ekkata yanawa"))

# Force a specific device
t = SinTransliterator(model="llm", contains_code_mix=True, device="cuda")
print(t.transliterate("mama office ekkata yanawa"))

Recommended on Kaggle, Google Colab, or any machine with a GPU for best performance. CPU inference is also fully supported.

---

Batch Transliteration

For converting large amounts of text, use transliterate_batch() instead of calling transliterate() in a loop. It feeds all inputs through the model in a single forward pass, making it significantly faster — especially on a GPU.

from sin_transliterator import SinTransliterator

t = SinTransliterator(model="transformer", contains_code_mix=False)

results = t.transliterate_batch([
    "mama yanawa",
    "kohomada",
    "mama iskole yanawa",
    "moko wenne"
])

for r in results:
    print(r)

For very large datasets, split into chunks to avoid running out of memory:

import pandas as pd
from sin_transliterator import SinTransliterator

df = pd.read_csv("your_file.csv")
t = SinTransliterator(model="transformer", contains_code_mix=False)

chunk_size = 32
results = []
for i in range(0, len(df), chunk_size):
    chunk = df["singlish"][i:i + chunk_size].tolist()
    results.extend(t.transliterate_batch(chunk))

df["sinhala"] = results
df.to_csv("output.csv", index=False)

---

Model Selection Guide

Scenario	model	contains_code_mix
Pure Sinhala, fast inference	"transformer"	False
Pure Sinhala, higher quality	"llm"	False
Sinhala + English mixed, fast	"transformer"	True
Sinhala + English mixed, higher quality	"llm"	True

When in doubt, start with model="transformer". It's faster, lighter, and performs well on the vast majority of standard Singlish input. Reach for model="llm" when output quality on complex or ambiguous inputs matters more than speed.

---

VRAM & Hardware Requirements

Both models are deliberately compact — well within reach of consumer hardware.

Model	Parameters	VRAM (estimated)	CPU Inference
Small100 (transformer)	~300M	~0.75 GB	Viable
Gemma3 (llm)	~300M	~0.60 GB	Viable

Any GPU with 2 GB+ VRAM handles both models comfortably. CPU inference is slower but fully functional — expect a few seconds per inference on a modern CPU.

---

API Reference

SinTransliterator

SinTransliterator(
    model="transformer",        # "transformer" | "llm"
    contains_code_mix=False,    # bool
    device=None,                # str — e.g. "cuda", "cpu", "mps"
    cache_dir=None              # str — custom cache path
)

Parameters

Parameter	Type	Default	Description
model	str	"transformer"	Architecture family. "transformer" uses the fine-tuned Small100 seq2seq model; "llm" uses the fine-tuned Gemma3 causal model.
contains_code_mix	bool	False	When True, loads the variant fine-tuned on Sinhala-English code-mixed data.
device	str	None	PyTorch device string. Auto-detected if not set ("cuda" if available, else "cpu").
cache_dir	str	None	Directory for caching downloaded weights. Defaults to ~/.cache/huggingface/.

.transliterate(text, max_new_tokens=256)

Parameter	Type	Default	Description
text	str	—	Romanised Sinhala or code-mixed input text.
max_new_tokens	int	256	Maximum number of tokens to generate.

Returns a str containing the Sinhala Unicode output.

.transliterate_batch(texts, max_new_tokens=256)

Parameter	Type	Default	Description
texts	list of str	—	List of romanised Sinhala or code-mixed input strings.
max_new_tokens	int	256	Maximum number of tokens to generate per input.

Returns a list of str containing Sinhala Unicode outputs in the same order as the input. Significantly faster than calling .transliterate() in a loop for large inputs.

---

Error Handling

from sin_transliterator import SinTransliterator
from sin_transliterator.exceptions import (
    InvalidModelError,
    ModelLoadError,
    TransliterationError,
)

try:
    t = SinTransliterator(model="transformer", contains_code_mix=False)
    result = t.transliterate("mama yanawa")
except InvalidModelError as e:
    print(f"Configuration error: {e}")
except ModelLoadError as e:
    print(f"Model load failed: {e}")
except TransliterationError as e:
    print(f"Inference error: {e}")

---

Acknowledgements

This package would not exist without the following contributions:

Deshan Sumanathilaka — for building and curating the core Singlish-Sinhala transliteration datasets used to train the base models.

• HuggingFace: deshanksuman
• Academic: [e-mail]
• Personal: [e-mail]

Rukshan Dias — for contributing the Sinhala-English code-mix dataset that made the code-mix model variants possible.

• LinkedIn: rukshan-dias
• Academic: [e-mail]

---

Contact

Savinu Linath Gunarathna

• HuggingFace: savinugunarathna
• Email: [e-mail]
• GitHub: savi664

---

Built with love for the Sinhala NLP community.