varuneshv/VCoder

1---2license: apache-2.03base_model: Qwen/Qwen2.5-Coder-3B-Instruct4pipeline_tag: text-generation5library_name: transformers6tags:7- code-generation8- python9- qwen10- unsloth11- transformers12- coding-assistant13language:14- en15---16 17# VCoder18 19VCoder is a Python-focused coding assistant fine-tuned from Qwen2.5-Coder-3B-Instruct using LoRA and Unsloth.20 21The model was trained on 15,000 Python instruction-response examples from the Python Code Instructions 15K dataset and optimized for Python code generation, problem solving, debugging, and algorithm implementation.22 23## Model Details24 25| Attribute | Value |26|------------|---------|27| Base Model | Qwen2.5-Coder-3B-Instruct |28| Fine-Tuning Method | LoRA |29| Framework | Unsloth |30| Dataset | Python Code Instructions 15K |31| Training Samples | 15,000 |32| GPU | NVIDIA Tesla T4 |33| Quantized Format | GGUF Q8_0 |34| Primary Language | Python |35 36---37 38## Training Pipeline39 40Training was performed incrementally:41 42| Stage | Samples |43|---------|---------|44| Stage 1 | 0 - 5,000 |45| Stage 2 | 5,000 - 10,000 |46| Stage 3 | 10,000 - 15,000 |47 48The model was trained using parameter-efficient fine-tuning (LoRA), allowing adaptation of the base model while keeping computational requirements low.49 50---51 52## Benchmark Results53 54![Output](https://cdn-uploads.huggingface.co/production/uploads/6a297050d3837ea7b12cc42f/BV8FY6fJN7KQ43jcpC6hr.png)55 56### HumanEval Comparison57 58The model was evaluated against the original Qwen2.5-Coder-3B-Instruct on HumanEval coding tasks.59 60| Model | Pass@1 |61|---------|---------|62| Base Qwen2.5-Coder-3B | 61.0% |63| VCoder | 68.0% |64 65### Improvement66 67```text68+7.0% Pass@1 improvement69```70 71This demonstrates that the fine-tuned model performs better on Python coding tasks than the original base model.72 73---74 75## Example Usage76 77### Python78 79```python80prompt = """81### Instruction:82Write a Python function to reverse a string.83 84### Input:85 86### Response:87"""88```89 90### Example Output91 92```python93def reverse_string(text):94    return text[::-1]95```96 97---98 99## Supported Tasks100 101- Python Code Generation102- Algorithm Design103- Data Structures104- Debugging105- Code Refactoring106- Coding Interview Questions107- Competitive Programming108- Function Completion109 110---111 112## GGUF Usage113 114Compatible with:115 116- Ollama117- LM Studio118- llama.cpp119 120---121 122## Training Dataset123 124Dataset used:125 126Python Code Instructions 15K127 128The dataset contains instruction-response pairs focused on Python programming tasks including:129 130- Function generation131- Data manipulation132- Algorithms133- Debugging134- Problem solving135 136---137 138## Limitations139 140- Primarily optimized for Python.141- Benchmark performed on a subset of HumanEval tasks.142- May generate incorrect code for highly specialized domains.143- Should not be used as the sole source of production-critical code.144 145---146 147## Acknowledgements148 149- Qwen Team for Qwen2.5-Coder150- Unsloth for efficient fine-tuning151- Hugging Face152- OpenAI HumanEval Benchmark153 154---155 156## Citation157 158```bibtex159@misc{vcoder2026,160  title={VCoder: Python Code Generation Model},161  author={Varunesh V, Prawin R K, Sarguru N},162  year={2026},163  base_model={Qwen2.5-Coder-3B-Instruct}164}165```166Github : https://github.com/varunesh-v167Mail : varunesh.wrk@gmail.com