Interactive explainer of how Gemma 4 E2B is trained: a four-stage pipeline from data preparation through pretraining with a Gemini teacher to instruction tuning, plus a hands-on demo where training steps sharpen a next-token prediction and lower the loss.

The pipeline
Training pipeline IngredientsTrillions of tokens, tokenizer PretrainingGuess the next token Gemini teacherShares its guesses Post-trainingFrom predictor to assistant gemma-4-E2B-itThe model you asked about
Click a stage for the plain-English version.
Pretraining, one step at a time
"The cat sat on the ___"  true next token: mat
mat ●24%
floor21%
chair20%
dog19%
moon16%
Steps: 0Loss: 1.43surprise at the right answer, lower is better