abubakar-siddik/bangla-alpaca

Bangla Alpaca

Bangla Alpaca is a culturally localized Bangla (বাংলা) adaptation of the Stanford Alpaca dataset. Unlike simple translation, this dataset uses native-first localization to produce natural, conversational Bangla instruction-following data for training high-quality LLMs.

📊 Overview

Aspect	Description
Language	Bangla (বাংলা)
Format	Instruction-Input-Output
Samples	~52K
License	Apache 2.0

📁 Dataset Structure

{
  "instruction": "একটি মেয়েকে নিয়ে একটি গল্প তৈরি করো যে একটি ভিনগ্রহে বেড়াতে যায়।",
  "input": "",
  "output": "একদা এক সময় হলি নামের এক তরুণী ছিল..."
}

• instruction: Task description in Bangla
• input: Optional context (can be empty)
• output: Expected response in Bangla

Usage

from datasets import load_dataset
from transformers import AutoTokenizer, Gemma3ForCausalLM, BitsAndBytesConfig
import torch

# Load dataset
dataset = load_dataset("abubakar-siddik/bangla-alpaca")

# Load model with 4-bit quantization
model_id = "google/gemma-3-1b-it"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

model = Gemma3ForCausalLM.from_pretrained(
    model_id,
    quantization_config=BitsAndBytesConfig(load_in_4bit=True),
    device_map="auto",
    torch_dtype=torch.float16,
)

🔖 Citation

@dataset{bangla_alpaca,
  title = {Bangla Alpaca: A Bangla Instruction Tuning Dataset},
  author = {Siddik, Abu Bakar},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/abubakar-siddik/bangla-alpaca}
}

📄 License

Apache 2.0