|
|
--- |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- python |
|
|
- code-generation |
|
|
- fine-tuned |
|
|
- qwen |
|
|
- gguf |
|
|
- coding |
|
|
- programming |
|
|
--- |
|
|
|
|
|
# Qwen2.5-Coder-3B-High |
|
|
|
|
|
[](https://opensource.org/licenses/Apache-2.0) |
|
|
|
|
|
> **Fine-tuned version of Qwen2.5-Coder-3B** optimized specifically for Python programming tasks. Outperforms the base model on Python-related problems, code generation, and real-world development scenarios. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🚀 Overview |
|
|
|
|
|
This repository hosts a fine-tuned variant of **Qwen2.5-Coder-3B**, trained on a high-quality dataset of Python programming problems, coding challenges, and real-world software engineering examples. The fine-tuning process significantly enhances the model’s ability to understand and generate idiomatic, efficient, and correct Python code. |
|
|
|
|
|
### ✅ Key Improvements Over Base Model: |
|
|
- Higher accuracy on Python syntax, standard library usage, and common frameworks (e.g., Pandas, NumPy, asyncio) |
|
|
- Better code completion and function generation from natural language prompts |
|
|
- Improved reasoning for algorithmic problems (e.g., sorting, recursion, data structures) |
|
|
- More consistent and readable output formatting |
|
|
|
|
|
--- |
|
|
|
|
|
## 📦 Model Files (GGUF Format) |
|
|
|
|
|
All models are provided in **GGUF** format for broad compatibility with inference engines like `llama.cpp`, `Ollama`, `LM Studio`, and more. |
|
|
|
|
|
| Filename | Quantization | Size | Recommended Use Case | |
|
|
|----------------------------------------|--------------|--------|-----------------------------| |
|
|
| `Qwen2.5-Coder-3B-High.F16.gguf` | Float16 | ~6.2 GB| Maximum quality (GPU) | |
|
|
| `Qwen2.5-Coder-3B-High.Q8_0.gguf` | Q8_0 | ~3.3 GB| High quality, CPU/GPU | |
|
|
| `Qwen2.5-Coder-3B-High.Q5_K_M.gguf` | Q5_K_M | ~2.2 GB| Balanced speed/quality | |
|
|
| `Qwen2.5-Coder-3B-High.Q4_K_M.gguf` | Q4_K_M | ~1.9 GB| Fast inference, low RAM | |
|
|
|
|
|
> 💡 **Recommendation**: Start with `Q5_K_M` for most local development tasks. |
|
|
|
|
|
--- |
|
|
|
|
|
## 📈 Performance |
|
|
|
|
|
Evaluated on an internal benchmark of 200 Python-specific prompts (including LeetCode-style problems, docstring-to-code, bug fixes, and library usage): |
|
|
|
|
|
| Metric | Base Qwen2.5-Coder-3B | Qwen2.5-Coder-3B-High | |
|
|
|--------------------------------|------------------------|------------------------| |
|
|
| Code Correctness (Pass@1) | 68% | **84%** | |
|
|
| Syntax Validity | 92% | **98%** | |
|
|
| Library Usage Accuracy | 71% | **89%** | |
|
|
| Readability (Human Eval) | 3.8 / 5 | **4.5 / 5** | |
|
|
|
|
|
> 🔍 *Benchmark details available upon request.* |
|
|
|
|
|
--- |
|
|
|
|
|
## 🛠️ Usage Examples |
|
|
|
|
|
### With `llama.cpp` |
|
|
```bash |
|
|
./main -m ./models/Qwen2.5-Coder-3B-High.Q5_K_M.gguf \ |
|
|
-p "Write a Python function that takes a list of integers and returns the sum of even numbers." \ |
|
|
-n 256 --temp 0.2 |
