Shaderwrap Finetune

Overview

Custom fine-tune of Qwen2.5-Coder (14B parameters) specialized for HLSL shader code generation. Curated a training dataset of hundreds of open-source and original URP shaders. Applied 5-bit quantization (~10GB) for local deployment via Ollama. The model powers ShaderWrap’s domain-specific code generation which translates natural language prompts into production-ready shader code with significantly improved accuracy over the base model.