Training Data

Definition

Training data is the data used to fit a model.

Simple form:

training examples -> parameter updates

The model uses training data to change its internal parameters.

Training data is what the model learns from directly.

---

What Training Data Contains

Training data can contain inputs alone or input-output pairs.

Examples:

input only:
text sequences

input-output pair:
image -> label
question -> answer
features -> target value

The exact structure depends on the learning setup.

Training data supplies the examples that drive learning.

---

Training Data Shapes The Model

The model is not learning in the abstract.

It learns from the examples it receives.

If the training data contains certain patterns, the model may learn them.

If important patterns are missing, the model may not learn them.

Training data defines much of the model's experience.

---

Training Data Is Not The Whole World

Training data is only a sample.

It may not cover every situation the model will face later.

Example:

training data:
examples already collected

future use:
new examples from the real environment

The gap between these matters for generalization.

A model learns from a sample, then gets judged on cases outside that sample.

---

Quality Matters

Training data quality affects what the model can learn.

Problems include:

incorrect labels
missing cases
biased sampling
duplicate examples
irrelevant noise

Bad training data can make a model learn bad patterns.

The model can only learn from the signal available in its training data.

---

Final Summary

• Training data is used to adjust model parameters.
• It can contain inputs or input-output pairs.
• It shapes what the model learns.
• It is only a sample of possible cases.
• Data quality strongly affects learned behavior.