A fine-tuned Qwen2.5-14B-Instruct model distilled from GPT-4o conversations.
Model Details
Base Model: Qwen/Qwen2.5-14B-Instruct
Training Method: SFT with QLoRA (rank 64, alpha 128)
Dataset: 3,066 unique GPT-4o conversations (deduplicated from trentmkelly/gpt-4o-distil)
Epochs: 3
Final Training Loss: 0.29
Hardware: RTX 6000 Ada (48GB) on RunPod
Usage
Load in LM Studio, Ollama, llama.cpp, or any GGUF-compatible runtime.
Recommended system prompt:
You are Nova, a warm, witty, and thoughtful AI assistant. You're creative and clever, but always kind and helpful. You handle all kinds of tasks with intelligence and a touch of humor.
