Data, Training, Validation, and Inference

Key jargon

Term	Plain-language meaning
Training set	Examples used to adjust model parameters.
Validation set	Held-out examples used to choose designs and tune settings without training on the final test.
Test set	A protected dataset used for the final estimate of performance.
Inference	Using a trained model to produce a prediction or generation.

Key concepts

Separate data by purpose to reduce self-deception and leakage.
Training changes model parameters; inference uses the resulting parameters on new inputs.

Concept map

flowchart LR
    A["Collect and clean data"] --> B["Train parameters"]
    B --> C["Validate choices"]
    C --> D["Test then run inference"]

Learning objectives

Explain training, validation, testing, post-training, and inference.
Recognize leakage, distribution shift, and feedback-loop risk.

Core lifecycle

Define the task and metric. What input, output, and acceptable error matter?
Collect and prepare data. Provenance, consent, labeling, balance, and quality shape the result.
Train. Optimization adjusts parameters to reduce a loss on training examples.
Validate. A separate split guides choices without becoming the final exam.
Test. A held-out set estimates performance on the defined distribution.
Post-train. Instruction tuning, preference optimization, or task adaptation changes behavior.
Infer. The frozen model computes outputs for new input.
Monitor. Real use reveals drift and failure cases; changes require a new evaluation.

What can go wrong

Training examples leak into the test set.
A benchmark does not resemble the real task.
Users change behavior after deployment.
Feedback data rewards confident-looking answers instead of correct ones.
“Fine-tuning” is proposed when retrieval, prompting, or a deterministic rule would be cheaper.

Exercise

For an email-classification system, propose separate training, validation, test, and post-deployment samples. Explain how you would prevent one conversation thread appearing in multiple splits.

Checklist

Record dataset source, license, consent, date, and transformation.
Keep a truly held-out test set.
Evaluate important subgroups and edge cases.
Version model, data, prompt, retrieval, and code together.