digitalai/DgMind-20B-Full

DgMind 20B: Advanced Reasoning & Expert Coding Assistant

DgMind 20B is a state-of-the-art, fine-tuned large language model designed for high-level logical reasoning and professional-grade software development. Built upon the GPT-OSS 20B architecture, this model has been optimized using the Unsloth library to provide efficient yet powerful performance on consumer-grade hardware.

👤 Identity & Developer

• Model Name: DgMind

• Developer: Erfan Mohamadnia

• Core Persona: A specialized AI assistant that excels in complex coding tasks, architectural decisions, and deep logical analysis.

📊 Training Details

• Base Model: GPT-OSS 20B (Unsloth 4-bit optimized)

• Dataset: Code-290k-ShareGPT

• Technique: LoRA (Low-Rank Adaptation)

• Optimization: Fine-tuned specifically on responses to enhance conversational accuracy and identity injection.

📈 Performance & Convergence

The model demonstrates a stable decrease in training loss, ensuring precise instruction following and a minimized hallucination rate in coding contexts.

💬 Prompt Template (Chat Format)

DgMind uses the following message structure to maintain context and role separation:

{% for message in messages %}{{ '<|start|>' + message['role'] + '<|message|>' + message['content'] + '<|end|>' }}{% endfor %}{% if add_generation_prompt %}{{ '<|start|>assistant<|message|>' }}{% endif %}

Example:

<|start|>user<|message|>Write a Python script for a custom API gateway.<|end|>

<|start|>assistant<|message|>

🛠 Deployment & Usage

Local Execution via Ollama

1. Download the .gguf file.

2. Create a file named Modelfile:

FROM "./DgMind-20B.Q4_K_M.gguf"

PARAMETER temperature 0.7

SYSTEM """You are DgMind, a helpful AI assistant developed by Erfan Mohamadnia. You specialize in advanced reasoning and expert-level coding."""

3. Run: ollama create DgMind -f Modelfile then ollama run DgMind.

Server Integration (llama.cpp)

Run the internal API server:

./llama-server -m DgMind-20B.Q4_K_M.gguf --host 0.0.0.0 --port 8080 --n-gpu-layers 62

📜 Acknowledgments

Special thanks to the Unsloth AI team for their memory-efficient fine-tuning kernels, and to ajibawa-2023 for providing the high-quality ShareGPT dataset.

---