Model Details

---

Training Pipeline

Training was performed incrementally:

The model was trained using parameter-efficient fine-tuning (LoRA), allowing adaptation of the base model while keeping computational requirements low.

---

Benchmark Results

HumanEval Comparison

The model was evaluated against the original Qwen2.5-Coder-3B-Instruct on HumanEval coding tasks.

Improvement

text
Copy code
+7.0% Pass@1 improvement

This demonstrates that the fine-tuned model performs better on Python coding tasks than the original base model.

---

Example Usage

Python

python
Copy code
prompt = """
### Instruction:
Write a Python function to reverse a string.

### Input:

### Response:
"""

Example Output

python
Copy code
def reverse_string(text):
    return text[::-1]

---

Supported Tasks

• Python Code Generation
• Algorithm Design
• Data Structures
• Debugging
• Code Refactoring
• Coding Interview Questions
• Competitive Programming
• Function Completion

---

GGUF Usage

Compatible with:

• Ollama
• LM Studio
• llama.cpp

---

Training Dataset

Dataset used:

Python Code Instructions 15K

The dataset contains instruction-response pairs focused on Python programming tasks including:

• Function generation
• Data manipulation
• Algorithms
• Debugging
• Problem solving

---

Limitations

• Primarily optimized for Python.
• Benchmark performed on a subset of HumanEval tasks.
• May generate incorrect code for highly specialized domains.
• Should not be used as the sole source of production-critical code.

---

Acknowledgements

• Qwen Team for Qwen2.5-Coder
• Unsloth for efficient fine-tuning
• Hugging Face
• OpenAI HumanEval Benchmark

---

Citation

bibtex
Copy code
@misc{vcoder2026,
  title={VCoder: Python Code Generation Model},
  author={Varunesh V, Prawin R K, Sarguru N},
  year={2026},
  base_model={Qwen2.5-Coder-3B-Instruct}
}

Github : https://github.com/varunesh-v Mail : varunesh.wrk@gmail.com