Gemma3-Code-Reasoning-4B-GGUF

This repository contains GGUF (GGML Universal Format) quantized versions of the GetSoloTech/Gemma3-Code-Reasoning-4B model, optimized for local inference with various quantization levels to balance performance and resource usage.

🎯 Model Overview

This is a LoRA-finetuned version of gemma-3-4b-it specifically optimized for competitive programming and code reasoning tasks. The model has been trained on the high-quality Code-Reasoning dataset to enhance its capabilities in solving complex programming problems with detailed reasoning.

🚀 Key Features

• Enhanced Code Reasoning: Specifically trained on competitive programming problems
• Thinking Capabilities: Inherits the advanced reasoning capabilities from the base model
• High-Quality Solutions: Trained on solutions with ≥85% test case pass rates
• Structured Output: Optimized for generating well-reasoned programming solutions
• Efficient Training: Uses LoRA adapters for efficient parameter updates
• Multiple Quantization Levels: Available in various GGUF formats for different hardware capabilities

📁 Available GGUF Models

Model File	Size	Quantization	Use Case
Gemma3-Code-Reasoning-4B.f16.gguf	7.77 GB	FP16	Highest quality, requires more VRAM
Gemma3-Code-Reasoning-4B.Q8_0.gguf	4.13 GB	Q8_0	High quality, good balance
Gemma3-Code-Reasoning-4B.Q6_K.gguf	3.19 GB	Q6_K	Good quality, moderate VRAM usage
Gemma3-Code-Reasoning-4B.Q5_K_M.gguf	2.83 GB	Q5_K_M	Balanced quality and size
Gemma3-Code-Reasoning-4B.Q4_K_M.gguf	2.49 GB	Q4_K_M	Good compression, reasonable quality
Gemma3-Code-Reasoning-4B.Q3_K_M.gguf	2.1 GB	Q3_K_M	Smaller size, moderate quality
Gemma3-Code-Reasoning-4B.Q2_K.gguf	1.73 GB	Q2_K	Smallest size, basic quality
Gemma3-Code-Reasoning-4B.IQ4_XS.gguf	2.28 GB	IQ4_XS	Intel optimized, good quality

🔧 Usage

Using with llama.cpp

# Download a GGUF model file
wget https://huggingface.co/GetSoloTech/Gemma3-Code-Reasoning-4B-GGUF/resolve/main/Gemma3-Code-Reasoning-4B.Q4_K_M.gguf

# Run inference with llama.cpp
./llama.cpp/main -m Gemma3-Code-Reasoning-4B.Q4_K_M.gguf -n 4096 --repeat_penalty 1.1 -p "You are an expert competitive programmer. Solve this problem: [YOUR_PROBLEM_HERE]"

Using with Python (llama-cpp-python)

from llama_cpp import Llama

# Load the model
llm = Llama(
    model_path="./Gemma3-Code-Reasoning-4B.Q4_K_M.gguf",
    n_ctx=4096,
    n_threads=4
)

# Prepare the prompt
prompt = """You are an expert competitive programmer. Read the problem and produce a correct, efficient solution. Include reasoning if helpful.

Problem: [YOUR_PROGRAMMING_PROBLEM_HERE]

Solution:"""

# Generate response
output = llm(
    prompt,
    max_tokens=4096,
    temperature=1.0,
    top_p=0.95,
    top_k=64,
    repeat_penalty=1.1
)

print(output['choices'][0]['text'])

🎛️ Recommended Settings

• Temperature: 1.0
• Top-p: 0.95
• Top-k: 64
• Max New Tokens: 4096 (adjust based on problem complexity)
• Repeat Penalty: 1.1

💻 Hardware Requirements

Quantization	Minimum VRAM	Recommended VRAM	CPU RAM
FP16	8 GB	12 GB	16 GB
Q8_0	5 GB	8 GB	12 GB
Q6_K	4 GB	6 GB	10 GB
Q5_K_M	3 GB	5 GB	8 GB
Q4_K_M	3 GB	4 GB	6 GB
Q3_K_M	2 GB	3 GB	4 GB
Q2_K	2 GB	2 GB	3 GB
IQ4_XS	3 GB	4 GB	6 GB

📈 Performance Expectations

This finetuned model is expected to show improved performance on:

• Competitive Programming Problems: Better understanding of problem constraints and requirements
• Code Generation: More accurate and efficient solutions
• Reasoning Quality: Enhanced step-by-step reasoning for complex problems
• Solution Completeness: More comprehensive solutions with proper edge case handling

🔗 Related Resources

• Base Model: GetSoloTech/Gemma3-Code-Reasoning-4B
• Training Dataset: GetSoloTech/Code-Reasoning
• Original Gemma Model: google/gemma-3-4b-it
• llama.cpp: GitHub Repository
• llama-cpp-python: PyPI Package

🤝 Contributing

This model was created using the Unsloth framework and the Code-Reasoning dataset. For questions about:

• The base model: Gemma3 Huggingface
• The training dataset: Code-Reasoning Repository
• The training framework: Unsloth Documentation

🙏 Acknowledgments

• Gemma Team for the excellent base model
• Unsloth Team for the efficient training framework
• NVIDIA Research for the original OpenCodeReasoning-2 dataset
• llama.cpp community for the GGUF format and tools

📞 Contact

For questions about this GGUF converted model, please open an issue in the repository.

---

Note: This model is specifically optimized for competitive programming and code reasoning tasks. Choose the appropriate quantization level based on your hardware capabilities and quality requirements.