Current Prefix
The model can only see these tokens right now, and its task is to predict "what is most likely to come next."
Training Progress0 / 6
Language Model Objective: Predicting the Next Token
The core task of training a large language model is not to "memorize entire answers," but to maximize the probability of the correct next token at each position. Think of it this way: given a prefix, the model repeatedly practices "what should come next" and uses loss to measure how far it is from the correct answer.
Observe next-token distribution
Watch loss decrease
Understand how sentences become training samples
Training Objective: Increase Probability of Correct Answers
First select a prefix scenario, then adjust the "Training Progress" slider. Higher training progress means higher probability assigned to the correct token, and corresponding lower loss. This is the core optimization direction of language model training.
What Loss Measures
If the model assigns most probability to the correct token, loss decreases; if it wastes probability on wrong candidates, loss remains high.
The Most Important Formula: Loss = -log P(correct token | current prefix). It focuses on one thing: whether the model has increased the probability of the correct answer.
One Complete Sentence Becomes Multiple Training Samples
During training, scoring doesn't happen only at the end of a sentence. Instead, a "prefix → next token" supervision signal is generated at every position. Click any position below to see what input and target the model actually faced.
Current Training Position
After seeing the prefix, the model needs to assign more probability to the actual next token.
Average Loss for This Sentence
A training sample typically contributes loss at multiple positions, which are then aggregated into the overall objective. Think of it this way: the model tries to make fewer mistakes at every step of the sentence.
Summary of this section: Language model training is not about "memorizing entire sentences," but breaking them into many next-token prediction problems and making the correct token more dominant at each step.
Key Takeaways After This Page
Training Objective is Simple
Essentially, it's about continuously increasing the probability of the correct next token, not understanding the "entire answer" at once.
Loss Reflects Probability
High loss usually means the correct token's probability is still too low; low loss means the model is closer to the correct distribution at that position.
Sentence Training Involves Multiple Positions
Each position in a sequence generates supervision signals, so the model learns at every step of the sequence, not just at the end.