Does inference data always lack labels?

No; labeled records can be replayed for evaluation, while labels remain separate from model inputs unless legitimately defined as inputs for another task.

Can two requests produce the same input?

Yes, after equivalent preparation.

Does inference automatically update model weights?

No; supplying an input need not perform training.

AI Glossary / Graphs and symbolic AI / Advanced

Inference Data

Definition

Inference data consists of the input values supplied to a trained machine-learning model when it produces a prediction or other output. Depending on the model, these inputs may be numerical features, encoded categories, text tokens, images, or another supported representation. The term describes the data's role in applying the model, rather than the data used to fit its parameters.

What it means

The source definition emphasizes input feature values. A prediction is the result of processing inference data, not the same thing as that input data.
“Inference” does not require a live request. A scheduled batch of records can also supply inference data. Nor must every record be something the model has never encountered: replaying an old example through a fixed model is still an inference operation.

How it works

An application obtains a request or scheduled record, retrieves any needed features, and transforms them into the model's expected input format. The model processes those values and returns its output.
The input contract needs to specify feature meaning, units, order or names, types, allowed categories, and treatment of missing values. Raw records and transformed model inputs can differ, so checking one does not necessarily validate the other.
Training and prediction should use compatible feature definitions and transformations. A mismatch can change a prediction even when the model itself is unchanged. Features must also be available at the intended prediction time; using information that arrives afterward can make historical evaluation misleading.
If a target label becomes available later, it can support evaluation. Its later availability does not make it a legitimate input to the earlier prediction.

Three practical examples

A delivery estimator receives distance, parcel weight, and a service-type code. The prepared feature values form its inference data; the predicted delivery duration is an output.
A document classifier receives a tokenized message. The original message and the numerical token sequence are different representations, and the model interface determines which representation it directly consumes.
A nightly inventory process supplies a batch of product records to a demand model. Batch execution changes when predictions are generated, but the records still function as inference data.

Think of it this way

Think of a calculator designed for meters receiving a length written in centimeters. The number is valid arithmetic input, yet its meaning is wrong for the calculation. A model can likewise accept a technically valid input whose units or interpretation differ from what it expects.

Worked example

Consider an invented model that estimates a parcel's handling time in minutes. Its input vector has the fixed order [weight_scaled, priority_code].
The recorded contract says weight_scaled equals weight in kilograms divided by 10. Priority_code is 1 for priority service and 0 for standard service. The model computes predicted_minutes = 4 + 6 × weight_scaled - 2 × priority_code.
A request describes a 5-kilogram priority parcel. Correct preparation gives [0.5, 1]. The model predicts 4 + 6 × 0.5 - 2 × 1 = 5 minutes. The vector [0.5, 1] is the inference data at this model boundary; 5 minutes is its output.
Now suppose the input record expresses weight as 5,000 grams. If the application converts grams to kilograms and then scales, it still supplies [0.5, 1] and obtains 5 minutes. Different raw representations can lead to identical inference data when preparation preserves their meaning.
If it skips the unit conversion and merely divides 5,000 by 10, it supplies [500, 1]. The same model now predicts 3,002 minutes. The arithmetic is consistent with the supplied values, but the feature preparation violates the contract.
Next, suppose the correct values are accidentally reversed to [1, 0.5]. The computed output becomes 9 minutes. A simple check that both values are numeric would miss this problem. The priority field's permitted values, its meaning, and the required vector order matter.
Finally, suppose priority status is missing and the system silently fills it with zero. The input [0.5, 0] produces 7 minutes. This calculation shows the effect of the chosen substitution; it does not establish that missing priority should mean standard service. That policy must be defined and evaluated rather than inferred from a convenient numerical default.
Assume the actual handling time is later observed to be 8 minutes. For the correctly prepared priority request, the absolute prediction error is |8 - 5| = 3 minutes. The observed 8-minute outcome can be joined to the earlier prediction for evaluation. It was not available as input before the parcel was handled.
All numbers and model coefficients here are invented to expose input-contract errors. They are not a validated handling-time model or evidence that any particular feature set is sufficient for real operations.

Illustrative case study

Fictional Willow Parcels introduces a second booking channel. Its existing channel records weights in kilograms, while the new channel records grams.
The prediction service continues responding successfully, so an availability dashboard shows no failure. However, duration estimates for the new channel become extreme. Willow compares raw records with the actual vectors sent to the model and finds the missing conversion.
The team adds checks for units, transformations, priority codes, and input ordering. It tests equivalent requests from both channels and confirms that they produce matching model inputs before judging prediction quality.

Common misconceptions and mistakes

“Inference data means predictions.” Predictions are outputs; this term concerns inputs.
“Accepted by the interface means suitable for the model.” Numeric validity does not verify meaning.
“Inference data must arrive in real time.” Batch inputs also qualify.
“Missing is equivalent to zero.” That depends on the feature's defined treatment.
“A known outcome can always be added as a feature.” It must have been available at the prediction time.

Business use

Assign ownership for each feature's meaning and preparation. When predictions change unexpectedly, compare the actual model inputs before assuming the learned model has deteriorated.

Check your understanding

Why do 5 kilograms and 5,000 grams yield the same correct vector?

Answer: Preparation converts their units consistently.

Why is [1, 0.5] invalid here?

Answer: The values occupy the wrong fields, and priority_code is not binary.

When can the observed 8 minutes be used?

Answer: After the outcome arrives, for evaluating the earlier prediction.

Related terms

Learn next

Prepare the correct input for a 12-kilogram standard parcel. Calculate its predicted duration, then identify the checks needed if the raw weight arrives as 12,000 grams.

Sources

Hopsworks MLOps Dictionary, Inference Data: https://www.hopsworks.ai/dictionary/inference-data Google Machine Learning Crash Course, monitoring pipelines and training-serving skew: https://developers.google.com/machine-learning/crash-course/production-ml-systems/monitoring

Previous: Inference attack Next: Inference engine