VCoder-Python-3B

VCoder - Python Code Generation Model

VCoder is a Python-focused coding assistant created by fine-tuning Qwen2.5-Coder-3B-Instruct using LoRA and Unsloth.

The goal of this project was to improve Python code generation capabilities while keeping training efficient enough to run on a single NVIDIA Tesla T4 GPU.

---

Project Overview

VCoder was trained on 15,000 Python instruction-response pairs from the Python Code Instructions dataset.

The model was fine-tuned incrementally in multiple stages and evaluated against the original Qwen2.5-Coder-3B-Instruct model using HumanEval coding benchmarks.

Objectives

• Improve Python code generation
• Enhance algorithm implementation
• Improve coding interview problem solving
• Maintain lightweight deployment through LoRA
• Support local inference using Ollama and GGUF

---

Tech Stack

Component	Technology
Base Model	Qwen2.5-Coder-3B-Instruct
Fine-Tuning Framework	Unsloth
Training Method	LoRA
Dataset	Python Code Instructions 15K
Training Environment	Google Colab T4
Evaluation	HumanEval
Deployment	Ollama
Quantization	GGUF Q8_0

---

Dataset

Dataset contains:

• Python coding tasks
• Function generation
• Algorithm implementations
• Debugging exercises
• Code explanations
• Programming problem solving

Training Samples Used:

15,000

---

Training Process

Training was performed in three stages.

Stage 1

Rows 0 - 5,000

Stage 2

Rows 5,000 - 10,000

Stage 3

Rows 10,000 - 15,000

This incremental approach reduced memory pressure and allowed training on limited GPU resources.

---

HumanEval Benchmark

VCoder was compared against the original base model using a 100-task HumanEval subset.

Model	Pass@1
Qwen2.5-Coder-3B-Instruct	61%
VCoder	68%

Improvement

+7 Pass@1 Points
+11.5% Relative Improvement

Benchmark Visualization

---

Example Usage

Ollama

FROM ./VCoder.Q8_0.gguf

Build:

ollama create vcoder -f Modelfile

Run:

ollama run vcoder

---

Sample Prompt

Write a Python function that returns the Fibonacci sequence up to n terms.

Sample Output

def fibonacci(n):
    sequence = []

    a, b = 0, 1

    for _ in range(n):
        sequence.append(a)
        a, b = b, a + b

    return sequence

---

Challenges Faced

During development several challenges were encountered:

• CUDA out-of-memory issues on T4 GPUs
• LoRA checkpoint management
• Sequence length optimization
• HumanEval benchmarking setup
• GGUF model conversion
• Ollama deployment

These issues were resolved through iterative experimentation and optimization.

---

Key Learnings

• Parameter-efficient fine-tuning using LoRA
• Efficient training with Unsloth
• Benchmarking LLMs using HumanEval
• Model quantization with GGUF
• Local LLM deployment using Ollama
• Hugging Face model publishing

---

Future Improvements

• Train on the bigger dataset
• Evaluate on complete HumanEval benchmark
• Add MBPP benchmark evaluation
• Expand support beyond Python
• Experiment with larger models
• Build a web-based coding assistant

---

Hugging Face

Model Repository:

https://huggingface.co/varuneshv/VCoder

---

Authors

Varunesh V, Prawin R K, Sarguru N

---

Acknowledgements

• Qwen Team
• Unsloth
• Hugging Face
• Ollama
• OpenAI HumanEval Benchmark

---

Mail : varunesh.wrk@gmail.com