FINE TUNE LAMMA USING QLORA

🦙 Fine-Tuning LLaMA-2-7B with QLoRA on Google Colab

This repository demonstrates how to fine-tune Meta’s LLaMA-2-7B model efficiently using QLoRA and PEFT on a single Google Colab GPU.
It turns the base model into a custom instruction-following chatbot trained on the Indian Constitution Articles dataset.

---

Check more models on Hugging Face

colab link : https://colab.research.google.com/drive/1-lxQKb9P8Xde0ft4hTLFFJ_O6V8vA8uo

📘 Dataset

• Name: nisaar/Articles_Constitution_3300_Instruction_Set
• Source: Hugging Face Datasets
• Description:
  Instruction-based dataset derived from the Indian Constitution, containing ~3300 examples.
  Each sample includes:

  {
    "instruction": "Explain Article 21 of the Indian Constitution.",
    "input": "",
    "output": "Article 21 guarantees the protection of life and personal liberty..."
  }
  

!pip install -q transformers accelerate bitsandbytes peft trl datasets einops 🧠 Model & Approach

• Base Model: meta-llama/Llama-2-7b-hf

• Quantization: 4-bit (using bitsandbytes)

• Finetuning Strategy: QLoRA (Quantized Low-Rank Adaptation)

• Trainer: SFTTrainer from trl

Key benefits:

• Runs on a single GPU (Colab T4 / A100)

• Significantly reduced VRAM usage

• Compatible with PEFT for adapter-based fine-tuning

Saving model : trainer.model.save_pretrained("Llama2-7b-constitution-qlora") tokenizer.save_pretrained("Llama2-7b-constitution-qlora")