Evaluation and data

What is a holdout set?

Data kept out of training and used only to measure performance.

What is the role of the final test set?

One unbiased estimate after every choice is made. Using it to tune anything makes it optimistic.

When is a random split appropriate?

When rows are independent and the future looks like the past, with no time order or repeated entities to worry about.

What is a temporal split?

Training on earlier data and evaluating on later data, which matches predicting the future.

What is an entity or group split?

Keeping all rows for the same user, customer, or patient on one side, so the model cannot score well by recognizing them.

What is distribution shift?

A difference between the data a model was trained or evaluated on and the data it meets in use.

What is covariate shift?

A change in the input distribution while the relationship between inputs and labels stays the same.

What is label shift (prior shift)?

A change in how common each class is, while what each class looks like stays the same.

What is concept shift?

A change in the relationship between inputs and labels, so the same input now has a different correct answer.

What is leakage in evaluation?

Information that would not exist at prediction time reaching the model, which inflates scores.

What is a temporal leak?

Using values recorded after the prediction time, such as today's totals joined to last year's rows.

What is a point-in-time join?

Joining each example only with feature values as they were known at that example's time.

How do duplicates cause leakage?

Identical or near-identical rows on both sides let the model score well by recall instead of generalization.

What is a label?

The correct answer attached to an example, used for training and scoring.

Why must a label have a precise definition?

'Churned' or 'urgent' can be computed many ways. Different definitions produce different models and scores.

What is label noise?

Errors or inconsistency in labels. It caps achievable accuracy and makes small score differences unreliable.

What are labeling guidelines?

Written rules telling labelers how to assign each label, with examples of hard cases.

What is inter-annotator agreement?

How often independent labelers give the same label. Cohen's kappa adjusts it for agreement by chance.

What is adjudication?

Resolving disagreements between labelers, usually by an expert or discussion against the guidelines.

What is a gold set?

A carefully labeled, checked set of examples used as the trusted reference for evaluation.

What is selection bias in evaluation data?

When the evaluated examples are not representative of real use, for example only cases a past system flagged.

What is a feedback loop in model data?

When a model's own decisions shape the data it is later trained or evaluated on, so it keeps confirming itself.

What is a slice in evaluation?

A subset of data, such as a region, device, or language, scored on its own to find weaknesses the average hides.

Why be careful with scores on small slices?

With few examples, the score has a wide margin of error. Report the count and an interval.

Why decide slices before looking at results?

Slices picked after seeing bad results will look bad partly by chance, which overstates the finding.

Why evaluate against a baseline?

A score only means something compared with a simple model, the current system, or human performance.

Why report a confidence interval with a metric?

It shows how much the score could move from sampling alone, so small differences are not over-read.

Why compare two models on the same examples?

A paired comparison cancels the shared difficulty of examples and detects real differences with fewer examples.

Why report metrics at the operating threshold?

That is where decisions are made. A good overall AUC can hide poor precision or recall at the chosen threshold.

Why does the positive rate affect precision?

With rarer positives, the same false positive rate produces more false alarms relative to true hits.

Why evaluate on data at the real class rate?

Scores on rebalanced or duplicated data do not reflect what happens in production.

Why must abstentions be counted in a metric?

A model that skips hard cases and is scored only on easy ones looks better than it is. Report coverage and accuracy together.

What is Goodhart's law?

When a measure becomes a target, it stops being a good measure, because people or models find ways to raise it without the real goal.

What is the difference between offline and online evaluation?

Offline uses held-out data. Online measures real users' behavior, usually in an A/B test.

What is a proxy metric?

A measurable stand-in for the real goal, such as clicks for satisfaction. It can move without the goal moving.

What are common ranking metrics?

Precision at k, recall at k, mean reciprocal rank (MRR), and normalized discounted cumulative gain (nDCG).

What does nDCG measure?

How well a ranked list puts the most relevant items near the top, with relevance discounted by position and normalized to the ideal order.

What does mean reciprocal rank measure?

The average of 1 divided by the rank of the first relevant result, across queries.

What is counterfactual (off-policy) evaluation?

Estimating how a new policy would perform from data logged under an old one, often by reweighting with the old policy's probabilities.

What is shadow evaluation?

Running a new model on live inputs and logging its outputs without acting on them, to compare it with the current model.

When is human evaluation needed?

When quality is subjective or open-ended, such as summaries or answers, and when checking that automated graders agree with people.

What is a grading rubric?

Clear criteria and examples for scoring outputs, so different graders score consistently.

What is a benchmark, and its limit?

A public, fixed evaluation set for comparing models. It may not match your task and may have leaked into training data.

Why version evaluation datasets?

If the set changes, old and new scores are not comparable, so a gain may just be an easier test.

What is a stale evaluation set?

One that no longer reflects current traffic, so it misses new kinds of inputs and errors.

What is error analysis?

Reading and grouping a sample of a model's mistakes to find their causes and what to fix next.

What is a datasheet or data card?

A document describing how a dataset was collected, labeled, and filtered, and what it should and should not be used for.

What does it mean for evaluation data to be representative?

Its mix of inputs, classes, and conditions matches what the model will see in use.

What evidence should a launch decision rest on?

Agreed primary and guardrail metrics on representative data, slice results, comparisons with the current system, and online results where possible.