File size: 3,165 Bytes
3855742 52c2ca9 3855742 52c2ca9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 | ---
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
pipeline_tag: text-generation
tags:
- code-generation
- python
- qwen
- unsloth
- gguf
- coding-assistant
language:
- en
---
# VCoder
VCoder is a Python-focused coding assistant fine-tuned from Qwen2.5-Coder-3B-Instruct using LoRA and Unsloth.
The 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.
## Model Details
| Attribute | Value |
|------------|---------|
| Base Model | Qwen2.5-Coder-3B-Instruct |
| Fine-Tuning Method | LoRA |
| Framework | Unsloth |
| Dataset | Python Code Instructions 15K |
| Training Samples | 15,000 |
| GPU | NVIDIA Tesla T4 |
| Quantized Format | GGUF Q8_0 |
| Primary Language | Python |
---
## Training Pipeline
Training was performed incrementally:
| Stage | Samples |
|---------|---------|
| Stage 1 | 0 - 5,000 |
| Stage 2 | 5,000 - 10,000 |
| Stage 3 | 10,000 - 15,000 |
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.
| Model | Pass@1 |
|---------|---------|
| Base Qwen2.5-Coder-3B | 61.0% |
| VCoder | 68.0% |
### Improvement
```text
+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
prompt = """
### Instruction:
Write a Python function to reverse a string.
### Input:
### Response:
"""
```
### Example Output
```python
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
### Ollama
```text
FROM ./VCoder.Q8_0.gguf
```
Build:
```bash
ollama create vcoder -f Modelfile
```
Run:
```bash
ollama run vcoder
```
---
## 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
@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 |