MicroGPT

A character-level GPT trained from scratch in Odin, running as WebAssembly.

0.70 


      

        About this demo
        
This is a ~119K parameter GPT-2-style transformer with 2 layers, 64-dimensional embeddings,
           and 4 attention heads. It was trained from scratch using a custom autograd engine written in Odin,
           then compiled to WebAssembly for inference. The entire model runs locally in your browser
           — no server calls.

        
Based on Andrej Karpathy's MicroGPT.