Gemma3 Tamil Translator

Model: arsath-sm/gemma3-tamil-translator

Fine-tuned Gemma-3 4B Instruct model specialized in high-quality English → Tamil translation.
Trained using Unsloth + LoRA on a single T4 GPU.

📊 Performance

Metric	Score	Note
BLEU	36.12	~25% relative improvement over base model
chrF++	65.25	Tested on 204-sentence held-out test set

This model significantly outperforms the base Gemma-3 4B and Llama-3 8B Instruct on English-to-Tamil translation tasks.

🚀 Model Details

Attribute	Value
Base Model	unsloth/gemma-3-4b-it-bnb-4bit
Architecture	Gemma-3 4B Instruct (8K context)
Method	LoRA (r=16, alpha=16) via Unsloth
Training Data	~1,836 high-quality English–Tamil parallel sentences
Training Time	~2.5 hours on Google Colab (T4 GPU)
Quantization	4-bit training, merged & saved in full precision

📦 Used Libraries & Requirements

To use this model, you will need the following libraries:

• transformers: For loading and running the model.
• torch: PyTorch backend.
• accelerate: (Optional) For optimized inference on GPU.

You can install them via pip:

pip install transformers torch accelerate

💻 Usage

Here is a simple example of how to use the model for translation:

from transformers import pipeline

# Initialize the translation pipeline
translator = pipeline(
    "text-generation",
    model="arsath-sm/gemma3-tamil-translator",
    device=0  # Use 0 for GPU, -1 for CPU
)

prompt = """You are a highly skilled translator. Translate the following English text to Tamil accurately and naturally.

English: Thank you so much for your help today."""

# Generate translation
output = translator(prompt, max_new_tokens=128, do_sample=False)[0]["generated_text"]

# Extract and print result
print(output.split("Tamil:")[-1].strip())
# Output: இன்று உங்கள் உதவிக்கு மிக்க நன்றி.

🎯 Intended Use

• Accurate English → Tamil translation
• Chat-style translation assistant
• Integration into Tamil NLP apps, chatbots, and education tools
• Research on low-resource Indic language translation

🤝 Attribution

Failed to fine-tune Gemma-3? Try Unsloth for 2x faster training and 0% memory degradation!
Model trained by Arsath.