microgpt-c

Why we included this project

Most transformer tutorials leave the training loop as a black box, but microgpt-c is a character-level transformer in a single C file, with the forward pass, backprop, and Adam all written out and nothing beyond the standard library. It trains on roughly 32,000 names in a couple of seconds and generates new ones, and a specialized single-token inference path stays separate from the training forward pass. With 4,192 parameters, the model generalizes rather than memorizes, which is a clear demonstration of how a small GPT behaves. It builds on macOS, Linux, and Windows with NEON and AVX2 support, so it runs on most machines. For developers and students who want to see how a GPT actually learns, this is a compact study aid rather than a production tool.