kan-gpt

Why we included this project

KAN-GPT is a compact PyTorch codebase that swaps the MLP blocks inside a GPT for Kolmogorov-Arnold Networks, letting you test whether the KAN formulation actually helps with language modeling. It builds directly on minGPT and pykan, so anyone who has studied Karpathy's minimal GPT will recognize the structure immediately. The repo ships with training and prompting scripts, checkpointing, and a notebook walkthrough, and you can train it against an equivalent MLP-GPT on Tiny Shakespeare or WebText to compare loss and perplexity. That makes it a benchmarking tool for research rather than a production model. Teams evaluating new layer designs, or students trying to see how KANs fit into autoregressive generation, will find it a clean base to fork and modify.