pirmber / transformer-from-first-principles

git clone https://github.com/pirmber/transformer-from-first-principles.git

A GPT-style Transformer built entirely from scratch with NumPy, featuring manual forward/backward propagation through attention, LayerNorm, softmax, residuals, GELU, and cross-entropy. Includes Adam, gradient checking, visualization, and autoregressive text generation—without autograd or deep-learning frameworks.