Add Training Script for Model Fine-tuning

[noor05-creator]

Problem

The repository currently only contains pre-trained LoRA adapter weights but no training code. This makes it impossible to reproduce the reported 41% GSM8K accuracy or experiment with different configurations.

Proposed Solution

I would like to contribute a complete training script (train.py) that implements:

• QLoRA fine-tuning on GSM8K dataset
• 4-bit quantization for memory efficiency
• Works on free Colab T4 GPU
• Matches existing adapter_config.json parameters
• Configurable command-line arguments

What I'll Add

• train.py - Complete training pipeline with QLoRA

Expected Outcome

Users can reproduce the 41% accuracy in ~1.5 hours on T4 GPU.

I would like to work on this under OScG'26.

@AshChadha-iitg Please assign this issue to me. Thank you!

[AshChadha-iitg]

@noor05-creator I’m assigning this issue to you under OSCG’26.

Please proceed with the standard workflow:

Fork this repository

Create a new branch (e.g., add-train-script)

Add a clean train.py that:

Uses QLoRA (4-bit)

Works on a free Colab T4

Matches adapter_config.json hyperparameters

Accepts basic CLI arguments (dataset path, epochs, lr, etc.)

Open a Pull Request referencing this issue.

If helpful, you may also add a short section to the README explaining how to run training.

I’ll review your PR once submitted , thanks for contributing!

[noor05-creator]

Alright! thank you