DhruvSIngh9911/Qwen-Adyapak-7B

Fine-Tuning Qwen2.5-Math-7B with LoRA (8-bit Quantization)

This repository demonstrates how to fine-tune the Qwen2.5-Math-7B model using LoRA in 8-bit precision for efficient memory usage. The process combines four CSV datasets (main_train.csv, main_test.csv, socratic_train.csv, socratic_test.csv) and produces a LoRA-adapted model for solving math/algebra problems.

---

Table of Contents

1. Overview
2. Requirements
3. Installation
4. Data Preparation
5. Fine-Tuning Steps
6. Inference
7. Merging LoRA Weights (Optional)
8. License

---

Overview

• Base Model: Qwen2.5-Math-7B
• Parameter-Efficient Fine-Tuning: LoRA (Low-Rank Adaptation)
• Quantization: 8-bit, using bitsandbytes
• Datasets: Four CSV files:
  • main_train.csv & main_test.csv
  • socratic_train.csv & socratic_test.csv

The code merges these training sets and fine-tunes the model to handle algebraic problem-solving tasks in a memory-efficient manner.

---

Requirements

• Python 3.8+
• PyTorch (tested with torch>=2.0.0)
• Transformers (tested with transformers>=4.30.0)
• PEFT (peft>=0.4.0)
• Datasets (datasets>=2.10.0)
• Accelerate (accelerate>=0.20.0)
• bitsandbytes (bitsandbytes>=0.39.0)
• pandas (for CSV reading)

---

Installation

git clone <this-repo-url>
cd <this-repo-folder>

# Install Python dependencies
pip install torch transformers datasets peft accelerate bitsandbytes pandas



## Data Preperation

- main_train.csv
- main_test.csv
- socratic_train.csv
- socratic_test.csv