Qwen2.5 Math Finetuning (NCERT Solutions)

This repository contains the code and configuration to finetune the Qwen2.5-0.5B-Instruct model specifically for solving NCERT-style mathematics problems. By leveraging LoRA (Low-Rank Adaptation), the model has been optimized to provide structured, step-by-step mathematical reasoning.

📊 Training Results & Metrics

The model showed significant convergence during the finetuning process. The drastic reduction in validation loss indicates that the model transitioned from general responses to highly specialized mathematical accuracy.

NOTE: As for some reason the jupyter file is not opening in github here is the link to original colab file. Link: https://colab.research.google.com/drive/1xCF9fCcP9cT1hw-peaTP-NlmiYiiTdmF?usp=sharing

Metric	Initial	Final	Improvement
Training Loss	0.0717	0.0440	~38.6% Reduction
Avg. Validation Loss	1.3250	0.0596	~95.5% Reduction

📈 Visualizing Convergence

Below are the curves captured during the sessions:

loss curves and log loss curves

ROUGE1, ROUGE1, ROUGE3, BLEU for the finetuned model

Validation loss before and after finetuning

---

🚀 Features

• Model: Qwen2.5-0.5B-Instruct
• Method: Parameter-Efficient Fine-Tuning (PEFT) using LoRA.
• Format: ChatML (System, User, Assistant).
• Task: Step-by-step NCERT Math Problem Solving.

---

📂 Project Structure

├── data/               # Training dataset (JSON)
├── src/                # Utility scripts for data & model loading
├── finetune.py         # Main training script
├── inference.py        # Test script for generating solutions
└── lora_adapters/      # Saved weights (generated after training)

🛠️ Installation & Usage

Install Dependencies

pip install -r requirements.txt

Start Finetuning

To run the training loop and save the adapters:

python finetune.py

Run Inference

To test the model on a new math problem:

python inference.py

📥 Quick Start: Using Pre-trained Adapters

If you wish to skip the training process, you can use the pre-trained LoRA adapters directly:

Extract the contents of lora_adapter.zip into a folder named lora_adapters/ in your project root.

Run Inference: The script will automatically load the base Qwen2.5-0.5B-Instruct model from Hugging Face and apply your local adapters.

---

🧪 Example Output

User: Find the roots of the equation $x^2 - 3x - 10 = 0$.

Assistant: To find the roots of the quadratic equation $x^2 - 3x - 10 = 0$, we can use the factorization method:

1. Identify two numbers that multiply to -10 and add to -3. These are -5 and +2.
2. Rewrite the equation: $(x - 5)(x + 2) = 0$.
3. Set each factor to zero: $x - 5 = 0$ or $x + 2 = 0$.
4. The roots are x = 5 and x = -2.

---

📜 License

This project is licensed under the MIT License.