---
language:
- ur
license: apache-2.0
base_model: unsloth/Llama-3.2-3B-Instruct
tags:
- urdu
- instruction-finetuning
- unsloth
- llama-3.2
- khurramcoder
datasets:
- large-traversaal/urdu-instruct
metrics:
- loss
model-index:
- name: Urdu-Llama-3.2-3B-Instruct-v1
  results: []
---

# Urdu-Llama-3.2-3B-Instruct-v1

Developed by **Khurram Pervez (Khurramcoder)**, this model is a fine-tuned version of Meta's Llama-3.2-3B-Instruct, specifically optimized for high-quality Urdu instruction following and generation.

## Model Highlights
- **Native Urdu Reasoning:** Trained on the `large-traversaal/urdu-instruct` dataset (51.7k rows), enabling the model to handle translation, creative writing, and QA tasks with cultural nuance.
- **Efficient Architecture:** Fine-tuned using **Unsloth** and QLoRA on an NVIDIA RTX 4060 Ti, making it a powerful yet lightweight 3B parameter model.
- **Optimized for 2026:** Uses the latest Llama 3.2 multilingual tokenizer for better Urdu script handling.

## How to Use
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Khurram123/Urdu-Llama-3.2-3B-Instruct-v1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")

instruction = "مصنوعی ذہانت کے مستقبل پر ایک مختصر نوٹ لکھیں۔"
prompt = f"### ہدایت:\n{instruction}\n\n### جواب:\n"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))