20s visual explainer sample
How an LLM Learns During Training
Training Tokens in Motion
Epoch
i
One pass over many text examples. More epochs give the model more chances to improve.
01
Loss
i
How wrong the predicted next token is compared with the target. Lower loss means less error.
0.82
Target prob
i
The probability assigned to the known target token. During training this should rise.
34%
Training progress
0%
Text examples tune model weights
LLM Training Loop
Prediction pass
Training tokens
Transformer layers
Target token
Training example
A neural network learns by [MASK]
Target token: adjusting
Teacher signal
Target: adjusting weights
Loss pushes probability toward target
Predict
Backprop
Update
Prediction forward
Backpropagation
Weights update
From wrong guesses to stronger next-token predictions
Example
Predict
Measure loss
Backprop
Update weights
Legend
LLM training terms and signals- Token activation
- A circle showing how a training token is represented inside the model.
- Model weight
- A learned connection strength that changes slowly during training.
- Prediction pass
- The model guesses the next token from the training text.
- Backpropagation
- Loss gradients flow backward through layers to calculate weight changes.
- Weight update
- Backprop nudges weights so useful paths strengthen and weak paths fade.
- Epoch, loss, target prob
- Epoch counts passes; loss should fall; target-token probability should rise.