Rebuilding GPT-style character models from scratch to learn PyTorch, autograd, and training loops.
A hands-on study repo following Andrej Karpathy’s makemore series. I re-implement character-level language models end-to-end: data preprocessing, n-gram/bigram baselines, MLPs with embeddings, and training loops with autograd—then iterate toward transformer-friendly patterns. The goal is to deeply understand how tokenization, batching, loss, and backprop fit together before scaling up.