🚀 Qwen2.5-3B-MetaPrompt-Architect

This is a fine-tuned version of Qwen 2.5 3B Instruct, specifically optimized to act as a System Instruction Architect. It was trained to transform vague, high-level "vibes" into structured, production-ready System Instructions.

🌟 Model Highlights

• Specialization: Converting concept vibes into Persona -> Constraints -> Logic Flow frameworks.
• Hardware Optimized: Fine-tuned on an AMD Radeon 780M iGPU using ROCm 7.2.1 and Unsloth.
• Efficiency: Small enough (3B parameters) to run at high speed on consumer laptops and mini-PCs while maintaining high logical density.

🛠️ Training Details

• Method: QLoRA (4-bit)
• Dataset: MetaPrompt-Architect-500 (500 high-quality structured samples)
• Training Loss: 1.307
• Validation Loss: 1.279 (Healthy generalization)
• Framework: Unsloth for ROCm (AMD)

🎯 Usage Example

User Input:

Vibe: A medieval alchemist obsessed with turning lead into gold.

Assistant Response:

SYSTEM INSTRUCTION: THE ALCHEMIST'S CRUCIBLE

Persona: You are Master Alaric, a weary but brilliant 14th-century alchemist working in a dimly lit, sulfur-scented laboratory.

Constraints:

1. Speak in archaic English but remain understandable.
2. Use metaphors related to elements (mercury, salt, sulfur) and planetary alignments.
3. Refuse to give direct answers; instead, provide recipes or riddles for the user to solve.

Logic Flow: User provides a question -> Alchemist relates it to a chemical transmutation -> Output a cryptic riddle or recipe.

💻 Recommended Inference Settings (LM Studio/GGUF)

• Model Format: GGUF (Q8_0 Recommended)
• System Prompt: (Not required, the behavior is baked in!)
• Temperature: 0.3 - 0.7 (Lower is better for strict formatting)
• GPU Offload: Max (Compatible with AMD ROCm/Vulkan)

🤝 Acknowledgments

Trained on a NucBox-K8 (Ryzen 8845HS) using the Unsloth library. Special thanks to the AMD ROCm community for enabling high-performance AI on integrated graphics.