Training Explorer
Step through LLM training in the browser
Description
The hardest part of learning the Transformer is that the formula is one line, yet nobody shows what the tensors actually do or which tokens a position can see. Training Explorer turns the whole pipeline into an experiment you can play, step and rewind: tokens, matrices, attention weights, residuals, FFN, logits and loss, all as real computed numbers.
The sample course uses a small configuration you can verify by hand: batch 1, 4 tokens, model dimension 4, 2 attention heads and a 6-word output vocabulary. Matrices and probabilities come from the same input in real time rather than a random heatmap.
Play and step:execute the model's internal computation event by event, with pause, rewind, jump and reset.
Follow tensor shapes:from [B,T] to [B,T,V], with every projection, head split and normalization marked on its dimension.
Chapter content:each chapter has derivations, numeric examples, Python and PyTorch code, advanced experiments, common mistakes and a quiz.
Event table and timing:maps animation frames to source, target, operation and tensor size.
Languages and themes:Simplified Chinese, English, light and dark.
Fully local:progress, language and theme live in your browser; no account and no backend.
The sample course uses a small configuration you can verify by hand: batch 1, 4 tokens, model dimension 4, 2 attention heads and a 6-word output vocabulary. Matrices and probabilities come from the same input in real time rather than a random heatmap.
Features
Play and step:execute the model's internal computation event by event, with pause, rewind, jump and reset.
Follow tensor shapes:from [B,T] to [B,T,V], with every projection, head split and normalization marked on its dimension.
Chapter content:each chapter has derivations, numeric examples, Python and PyTorch code, advanced experiments, common mistakes and a quiz.
Event table and timing:maps animation frames to source, target, operation and tensor size.
Languages and themes:Simplified Chinese, English, light and dark.
Fully local:progress, language and theme live in your browser; no account and no backend.
