🔧 Fine-tuning LLMs for Python Code Generation

This project focuses on fine-tuning Microsoft's phi-1.5 language model for Python code generation using natural language problem descriptions. It leverages the CodeParrot/APPS dataset and utilizes efficient fine-tuning techniques like PEFT and LoRA to achieve significant performance improvements on code synthesis tasks.

---

📌 Project Overview

• Goal: Enhance Python code generation capabilities of a pretrained LLM using domain-specific fine-tuning.
• Base Model: Microsoft phi-1.5 (a small yet effective LLM)
• Dataset: CodeParrot / APPS (10,000 Python problems with solutions)
• Techniques Used: PEFT (Parameter-Efficient Fine-Tuning), LoRA (Low-Rank Adaptation)
• Environment: Google Colab (GPU-enabled)

---

📁 Repository Contents

This repository contains the following files:

1. Fine-Tuning LLM Tasks.ipynb: Jupyter Notebook with the code and implementation for fine-tuning the language model (LLM) for code generation tasks. You can open this notebook to explore the step-by-step code and results.

2. Fine_tune_LLM_for_Code_Generation.pdf: A detailed & complete documentation file that explains the methodology, techniques, and findings from the fine-tuning process. It provides in-depth insights into the code generation improvements achieved through fine-tuning.

3. Datasets: The dataset folder includes the following JSON files:

• trainDataset.json: The training dataset used for fine-tuning the model for first time.
• SyntheticDataset.json: Includes the synthetic dataset generated using LLM for fine-tuning the model for second time.
• CombinedDataset.json: Contains the combined dataset (train + synthetic) used for final fine-tuning.
• testDataset.json: The test dataset used for evaluating the model's performance.

---

📁 Dataset Overview

• Source: CodeParrot / APPS
• Description: A collection of English programming problem statements paired with corresponding Python solutions.
• Dataset Used: First 500 samples.
  • Training: 300 samples
  • Validation: 100 samples
  • Testing: 100 samples

---

🛠️ Tools & Libraries

• Python – Core language used for training and scripting.
• Google Colab – Platform for training with free GPU resources.
• Hugging Face Transformers – Loading and interacting with pretrained models.
• PEFT Library – Implements LoRA for efficient model fine-tuning.
• Datasets – Used to load and process the CodeParrot/APPS dataset.
• Accelerate – Facilitates training on available hardware (Colab GPU).
• ROUGE Evaluation – Used to evaluate generated Python code quality.

---

🔁 Methodology

1. Preprocessing

• Filter and clean dataset
• Tokenize using phi-1.5 tokenizer
• Split into training, validation, and test sets

2. Fine-Tuning Strategies

• Original Data Training: Fine-tune using human-annotated problems
• Synthetic Data Training: Generate synthetic problem-code pairs using ChatGPT and fine-tune
• Combined Data Training: Merge both datasets and perform final fine-tuning

3. Training Setup

• Used LoRA to fine-tune efficiently on limited hardware
• Conducted training on Google Colab with ~16GB GPU memory

---

📈 Evaluation

• Metrics: ROUGE (Recall-Oriented Understudy for Gisting Evaluation)
  • ROUGE-1: Unigram overlap
  • ROUGE-2: Bigram overlap
  • ROUGE-L: Longest common subsequence
• Results:
  • Best performance observed when trained on combined dataset
  • Achieved ~30% improvement in code generation quality (based on ROUGE scores)

---

🧠 Key Learnings

• Domain-specific fine-tuning greatly enhances LLM performance for code generation.
• Synthetic data generated using OPEN AI API helps boost generalization.
• LoRA + PEFT allows efficient training even on resource-limited platforms like Google Colab.

---

🚀 Future Improvements

• Expand to support multiple programming languages.
• Evaluate using additional metrics like BLEU, CodeBLEU, and Human Eval.
• Explore model deployment as a code-assist tool or IDE extension.

---

📎 Acknowledgments

• Microsoft's phi-1.5 Model
• CodeParrot/APPS Dataset
• Hugging Face Transformers and PEFT libraries

---

📬 Contact

If you have any questions or would like to collaborate, feel free to reach out:

Prateek Kumar Sharma
📧 Prateeksharma0112@gmail.com
🌐 LinkedIn