Worked example
The linear layer
Live workbench
Neural Network
Epoch0
Train loss—
Val loss—
Examples/sec0
Ready
Live workbench
Token Workbench
Tokens0
Vocabulary0
Transitions0
Perplexity—
Each chip is one token. Click any token to see what tends to follow it.
After the token
—these are most likely next:
Press Generate to sample a passage from the model.
Updates0
Epoch0.00
Loss—
Chars/sec0
Params0
Live sample
Press Train to start. A tiny one-block transformer learns character by character on the corpus above; its output drifts from noise toward Shakespearean text as the loss falls. Then Sample to read what it has learned.
Updates0
Loss—
Vector analogy
−
+
Nearest tokens
- Train, then Solve.
Honest note. These vectors come from the model you trained on Shakespeare; classical directions (royalty, capital) only partly emerge at this scale — more training and data sharpen them.