GPT-2, stage by stage
Type a sentence and follow it through every stage of GPT-2 small (the 124-million-parameter model of 2019): words → tokens → vectors → block 1, …, block 12 → output layer → loss. The model runs inside this page (its weights, 167 MB, are downloaded once and then cached by your browser), so every number you see is computed from your sentence. The diagram on the left shows where in the architecture you are, and the panel on the right lists the hyperparameters that matter at that stage. Companion page for the MATH 289 lectures on LLM architecture.
try:
loading …