Sinhala-Qwen3-v7500

Sinhala Qwen 3 - Fine-Tuned Model (v7500)

Model Description

This is a fine-tuned version of the Qwen 3 base model, specifically trained to understand and generate the Sinhala language.

Originally, the base model had absolutely no understanding of Sinhala. Through custom fine-tuning using the Unsloth library, this model has been taught the foundational elements of the language from scratch. It is now capable of basic Sinhala comprehension and text generation, marking a significant step in low-resource language adaptation for this architecture.

Model Details

• Base Model: Qwen 3 (Qwen3ForCausalLM)
• Language(s): Sinhala (si), English (en)
• Fine-Tuning Library: Unsloth (unsloth_version: 2026.1.4)
• Checkpoint/Version: v7500 (Trained up to 7500 steps)
• Status: Experimental / Early Stage

Intended Use

This model is intended for researchers and developers working on Natural Language Processing (NLP) for the Sinhala language. It can be used as a starting point for further fine-tuning, vocabulary expansion, or basic Sinhala text generation tasks.

Note: Since the model was trained from scratch to learn a completely new language, it might still make grammatical errors or occasionally struggle with complex sentence structures. It represents a foundation that is continuously being improved.

How to Use

You can use this model directly with the transformers library in Python:

from transformers import AutoModelForCausalLM, AutoTokenizer


model_id = "sh4lu-z/Sinhala-Qwen3-v7500" 

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "ඔබට කොහොමද?"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=50)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Procedure

This model was trained using the Unsloth library to significantly speed up the fine-tuning process while maintaining accuracy.

Framework: PyTorch / Transformers

Hardware: Fine-tuned on GPU instances