metindeder/test-master-unit-test-model-GGUF

TestMaster-7B-GGUF: Resource-Efficient Polyglot Unit Test Generator

TestMaster-7B is a specialized Large Language Model (LLM) fine-tuned for generating robust, industrial-grade unit tests across multiple programming languages. Built upon Qwen2.5-Coder-7B-Instruct and fine-tuned using Unsloth, this model is optimized for logic reasoning, edge-case detection, and mock architecture simulation.

This repository contains the 4-bit Quantized (GGUF - Q4_K_M) version, designed to run efficiently on consumer hardware (e.g., NVIDIA RTX 3070, 8GB VRAM) with minimal performance loss compared to the FP16 baseline.

🚀 Key Features

• —Resource Efficient: Runs on ~4.5 GB of VRAM/RAM.
• —Polyglot Capabilities: Master-level performance in C, Java, Go, Python, C++, JavaScript, and even ROS2.
• —Advanced Reasoning: Capable of handling complex scenarios like:
• —Memory Safety (Null pointer checks in C).
• —Concurrency (Goroutines/WaitGroups in Go).
• —Reflection & Retry Logic (Java).
• —Async/Promise Mocking (JavaScript/Jest).
• —ASTER Methodology: Trained with principles inspired by Automated Software Testing & Error Remediation, focusing on compilability and behavioral correctness.

📊 Performance Evaluation

The model has been rigorously tested across various programming languages and complex testing scenarios. Below is the updated summary of its performance:

<details> <summary>📋 Click to Show All Training Data (Log)</summary>

</details>

💻 Usage

Prompt Format (ChatML)

This model uses the ChatML template. Strict adherence to this format is recommended for optimal results.

<|im_start|>system
You are an expert software tester. Your goal is to write a comprehensive unit test.<|im_end|>
<|im_start|>user
Instruction:
Write a unit test for the following [Language] code...

Input:
[Source Code Here]
<|im_end|>
<|im_start|>assistant

Running with LM Studio / Ollama

1. 1.Download the .gguf file from this repository.
2. 2.Load it into your preferred GGUF runner (LM Studio, Ollama, etc.).
3. 3.System Prompt: Set the system prompt to: "You are an expert programmer and unit test generator."
4. 4.Context Window: Recommended set to 4096 or 8192.

Running with Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="./unit-test-model-q4_k_m.gguf",
    n_ctx=4096,
    n_gpu_layers=-1, # Offload all layers to GPU
    verbose=False
)

prompt = """<|im_start|>system
You are an expert unit test generator.<|im_end|>
<|im_start|>user
Write a Python unit test for a function that divides two numbers.<|im_end|>
<|im_start|>assistant"""

output = llm(
    prompt,
    max_tokens=1024,
    stop=["<|im_end|>"],
    echo=False
)

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

⚠️ Limitations & Bias Python Imports: In very complex Python scenarios involving obscure libraries, the model might occasionally miss an import statement or hallucinate a mock path.

Context Window: While optimized, extremely large source code files might exceed the context window. It is recommended to test modular functions or classes.

Self-Healing: For dynamic languages (Python, JS), we recommend using this model in a "Self-Healing" loop (Execution -> Error Capture -> Repair) for 100% reliability.

📜 License This model is a fine-tune of Qwen2.5-Coder and is licensed under Apache 2.0.

🤝 Acknowledgments Fine-tuned using Unsloth (2x faster training).

Base model by Qwen Team.

Dataset curated for ASTER (Automated Software Testing & Error Remediation) research.