What changes when a language model learns?

An assistant uses a new name correctly after you introduce yourself. Did its model weights just change? Usually, using information in the current conversation is different from training the underlying model. To see the difference, follow the data through one training step.

Before you begin: Understand next-token generation, parameters, and loss.

Create a target from text

Take the sentence “The class starts at noon.” In a simplified next-token task, the model sees a prefix and predicts the next token. Text itself supplies the target. This is self-supervised learning: the training procedure constructs labels from available data instead of asking a person to label every prediction.

Real tokenization may split the sentence differently from its visible words. Training examples also involve formatting, length limits, and sometimes masks that decide which positions contribute to the loss. The simplified sentence is enough to follow the objective without pretending to describe every training implementation.

Measure a probability, not just the final guess

Suppose the correct target token is “noon.” One model assigns it probability 0.8; another assigns it 0.2. Both might occasionally sample the correct word, but the first gives that target more probability. A common loss is negative log probability: loss = -log(p), where p is the assigned probability of the observed target and log is the natural logarithm.

For 0.8, the loss is about 0.223. For 0.2, it is about 1.609. Lower is better for this prediction objective. The logarithm penalizes assigning very little probability to what actually occurred. This is a mathematical score for the data, not a direct measure of truthfulness or helpfulness.

Update the model carefully

Backpropagation calculates gradients, which describe how a small parameter change would affect the loss. An optimizer uses them to change the parameters. The learning rate controls the size of an update. If steps are poorly chosen, learning may be unstable or unnecessarily slow.

Training typically uses batches of examples. An epoch is one pass through a dataset under the chosen sampling setup. Large-model training is often planned in tokens and optimization steps, so a neat whole number of epochs is not always the useful unit.

The model sees many contexts, including imperfect and conflicting text. Becoming better at predicting that text can teach useful structure while also absorbing stereotypes, mistakes, or memorized sequences. Data preparation and evaluation are part of training, not optional cleanup afterward.

Test beyond the practice material

Training loss answers how well the model fits its training objective on the material used to update it. Validation loss checks held-out material. Task evaluations ask more specific questions: does the model follow instructions, preserve quoted dates, answer in the right language, or identify missing evidence?

A model can improve on an average score while getting worse on a small group of important examples. That is why later lessons separate aggregate metrics from inspection of particular failures. Also check whether evaluation material was exposed during training or tuning; leaked answers make progress difficult to interpret.

Conversation context is another mechanism

When you write “My project is called Cedar,” the current prompt can include that fact on later turns. The model can use it without any parameter update. This is often called in-context learning when examples or instructions in context shape behavior.

An application's saved memory is a third mechanism: the product may store information and insert it into later requests. None of these mechanisms should be assumed from the assistant saying “I remember.” Data use for future provider training is a separate policy question that depends on the service and account controls.

Calculate the loss yourself

This complete Python 3 example evaluates two invented predictions. It does not train or contact a model.

# Runnable: Python 3, standard library.
import math

for probability in [0.8, 0.2]:
    print(round(-math.log(probability), 3))

Try a probability of 0.01. Predict whether the loss should rise or fall. Then explain why a model with lower next-token loss might still answer the room-booking question incorrectly.

Next, we will examine the finite space available for the conversation itself. That explains why a model can use a fact on one turn and appear to lose it later.

Further reading

Training language models to follow instructions with human feedback describes one post-training pipeline. Language Models are Few-Shot Learners investigates behavior shaped by examples supplied in context.

Continue to the next lesson.