How many Refer to the Exhibit Practice Questions questions are on this page?

This page has 11 Refer to the Exhibit Practice Questions scenario questions for the AI0-001 exam, each with detailed explanations and wrong-answer analysis.

How should I approach AI0-001 scenario questions?

Read the full scenario before looking at the answer options. Identify the constraint or requirement in the scenario, then eliminate options that are generally true but wrong for this specific case. Scenario questions reward careful reading over pattern matching.

← Back to CompTIA AI+ AI0-001 questions

Scenario-based practice

Refer to the Exhibit Practice Questions

Practise CompTIA AI+ AI0-001 practice questions — original exam-style scenarios covering every exam domain, with detailed explanations, wrong-answer analysis, and common exam traps.

Start full practice test Read exam guide

scenario questions

AI0-001

exam code

CompTIA

vendor

Scenario guide

How to approach refer to the exhibit practice questions

Practise exhibit-style questions that ask you to read a topology, table, command output or diagram before choosing the best answer.

Quick answer

Exhibit-style questions test whether you can read a topology, command output, diagram or table before choosing the best answer.

How to extract the relevant detail from an exhibit.

How topology, command output or routing information affects the answer.

How to avoid answering from memory before reading the evidence.

How to map the exhibit back to the exam objective.

Practice scenarios

Question 1mediummultiple choice

Full question →

Refer to the exhibit. A data engineer runs a validation report on the customers table. The "income" column has 12 null values. Which imputation strategy is most appropriate for this column?

Exhibit

Data Validation Report:
Table: customers
- column "age": null values: 0, unique values: 87, min:18, max:99
- column "income": null values: 12, unique values: 1500, min:0, max:500000
- column "region": null values: 0, unique values: 4, values: ["North", "South", "East", "West"]
- column "gender": null values: 0, unique values: 2, values: ["M", "F"]

A
Remove rows with null income
Why it fails: Deleting rows discards the remaining valid attributes of those 12 customers, biasing any downstream analysis. Row removal suits columns with very few nulls where the lost records are negligible and no meaningful distribution shift results; here imputation preserves the data.
B
Replace nulls with the median income per region
Income is typically right-skewed and contains outliers, so the mean would be distorted; the median is robust. Computing it per region preserves local wage differences, satisfying the requirement to impute the 12 nulls without flattening genuine geographic variation in the column.
C
Replace nulls with 0
Why it fails: Zero income is unrealistic and would introduce bias.
D
Replace nulls with the mean income of the entire dataset
Why it fails: Mean imputation assumes a symmetric distribution; the exhibit shows income is skewed, so the mean is pulled by high earners and distorts the 12 replacements. It is tempting because mean imputation is the standard default for roughly normal numeric columns, where it would indeed be the correct choice.

Full breakdown with real-world context →

Question 2mediummultiple choice

Full question →

Refer to the exhibit. An AI auditor reviews the fairness configuration. What is the purpose of this policy?

Exhibit

{
  "fairness_metric": "demographic_parity",
  "threshold": 0.1,
  "protected_attributes": ["race", "gender"]
}

A
Ensure equal error rates across groups
Why it fails: Equal error rates is a genuine fairness objective, but the exhibit's policy targets the specific fairness metric shown, not error-rate parity across groups. It is tempting because equalised odds and equal opportunity are standard fairness definitions, and would be correct if the configuration measured false positive or false negative rates per group.
B
Ensure equal positive prediction rates across groups
Equal positive prediction rates across groups is the definition of demographic parity, so the policy constrains the model to produce the same proportion of favourable outcomes for each protected group, regardless of differing base rates or accuracy.
C
Ensure equal accuracy across groups
Why it fails: Equal accuracy across groups measures overall correctness, but fairness policies typically equalise error rates or outcomes rather than raw accuracy, which can mask unequal false positive and false negative distributions. It is tempting because accuracy parity sounds like fairness, and would be right if the policy explicitly targeted predictive performance parity.
D
Ensure model interpretability
Why it fails: Interpretability explains how a model reaches decisions, whereas a fairness configuration targets disparate impact across protected groups. It is tempting because interpretability tooling is often deployed alongside fairness dashboards, and would be correct if the policy's goal were explaining individual predictions rather than measuring group outcomes.

Full breakdown with real-world context →

Question 3easymultiple choice

Full question →

Refer to the exhibit. What is the recall of the model?

Exhibit

Predicted Negative   Predicted Positive
Actual Negative      9000                 100
Actual Positive       500                 400

A
0.44
Recall equals true positives divided by the sum of true positives and false negatives. Reading the confusion matrix, 0.44 results from that ratio, giving the proportion of actual positives the model correctly identified. This satisfies the stem's request for the recall value derived from the exhibit.
B
0.80
Why it fails: 0.80 is the precision (TP/(TP+FP) = 400/500 = 0.8).
C
0.90
Why it fails: 0.90 is the specificity (TN/(TN+FP) = 9000/9100 ≈ 0.989, not 0.90).
D
0.99
Why it fails: 0.99 is the accuracy ((9000+400)/10000 = 0.94, not 0.99).

Full breakdown with real-world context →

Question 4hardmultiple choice

Full question →

Refer to the exhibit. A team deploys a sentiment analysis model with this policy. After one month, the monitoring system triggers an alert for feature drift. Which action should the team take first?

Exhibit

Refer to the exhibit.

JSON Policy for Model Deployment:
{
  "model": "sentiment_analysis_v2",
  "threshold": 0.7,
  "fairness_check": {
    "protected_attributes": ["gender", "age_group"],
    "metric": "demographic_parity",
    "tolerance": 0.05
  },
  "explainability": {
    "method": "LIME",
    "num_features": 5
  },
  "monitoring": {
    "drift_detection": {
      "feature_drift": true,
      "prediction_drift": true,
      "alert_threshold": 0.2
    }
  }
}

A
Review the fairness check settings to ensure protected attributes are still relevant.
Why it fails: Feature drift concerns input data distributions shifting, not protected-attribute fairness, so fairness settings cannot explain or resolve the alert. It is tempting because fairness monitoring is part of the same policy, and would be the right first step for a bias or disparate-impact alert.
B
Immediately retrain the model on recent data to adapt to the drift.
Why it fails: Retraining immediately treats the symptom without diagnosing the cause; drift alerts require investigating data quality, pipeline changes or upstream schema shifts first, since retraining on corrupted recent data bakes in the fault. Retraining is the right response once drift is confirmed as genuine concept change rather than a data-collection defect.
C
Compare the current feature distributions with the training set to identify which features drifted.
Feature drift means input distributions have shifted away from the training data, so the first step is diagnostic: compare current feature distributions against the training set to identify which features drifted and by how much, before deciding whether to retrain or adjust monitoring thresholds.
D
Reduce the classification threshold to 0.5 to increase sensitivity.
Why it fails: Lowering the threshold alters prediction sensitivity, not the drifted input features, so it masks rather than addresses the drift. It is tempting because threshold tuning is a quick lever for shifting precision and recall, and would be correct when the alert concerns class imbalance or recall degradation.

Full breakdown with real-world context →

Question 5mediummultiple choice

Full question →

Refer to the exhibit. A machine learning pipeline configuration is shown. During a deployment, the model evaluation passes with accuracy 0.86 and precision 0.79. However, the pipeline proceeds to deploy. What is the most likely reason for this behavior?

Exhibit

Refer to the exhibit.

```
{
  "pipeline_version": "2.1",
  "components": {
    "data_ingestion": {
      "source": "s3://data-bucket/transactions/",
      "schedule": "cron(0 2 * * ? *)"
    },
    "feature_engineering": {
      "script": "features.py",
      "parameters": {
        "window_size": 7,
        "aggregation": "mean"
      }
    },
    "model_training": {
      "algorithm": "xgboost",
      "hyperparameters": {
        "n_estimators": 100,
        "learning_rate": 0.1
      },
      "training_data_version": "v1"
    },
    "model_evaluation": {
      "metrics": ["accuracy", "precision", "recall"],
      "threshold": {"accuracy": 0.85, "precision": 0.80}
    },
    "model_deployment": {
      "target": "production",
      "rollback_condition": "if_accuracy_drops_below_0.85"
    }
  }
}
```

A
The precision metric is not included in the evaluation script
Why it fails: Precision is reported as 0.79, so it is clearly computed; the gate simply compares accuracy against its threshold and ignores precision. It is tempting because missing metrics do cause silent passes, and would be the correct diagnosis if the exhibit showed no precision value at all.
B
The deployment only checks the accuracy threshold for rollback condition
The pipeline gates deployment solely on the accuracy threshold, so a 0.86 accuracy score passes even though precision of 0.79 falls below its own gate. The rollback condition never evaluates precision, allowing deployment despite the weaker metric.
C
The deployment target is set to staging instead of production
Why it fails: The deployment target determines where artefacts land, not whether the evaluation gate passes; the pipeline would proceed regardless of staging or production. It is tempting because staging deployments are lower risk, and would be the right answer if the question asked why a model reached staging rather than production.
D
The operator manually overrode the threshold
Why it fails: Nothing in the exhibit indicates a manual override, and the pipeline proceeded automatically on its own gate result. It is tempting because overrides do bypass gates in practise, and would be correct if logs showed an operator approving the run despite a failed threshold check.

Full breakdown with real-world context →

Question 6easymultiple choice

Full question →

Refer to the exhibit. A data engineer is training a binary classification neural network. The loss fluctuates and does not converge. Which hyperparameter adjustment is most likely to stabilize training?

Exhibit

model:
  type: Sequential
  layers:
    - type: Dense
      units: 128
      activation: relu
    - type: Dense
      units: 64
      activation: relu
    - type: Dense
      units: 1
      activation: sigmoid
optimizer:
  type: Adam
  learning_rate: 0.01

A
Change the activation to tanh
Why it fails: Switching to tanh changes the activation function's output range but does nothing to correct the oversized weight updates driving the divergence. Activation choice addresses vanishing gradients or output range, not oscillation. Tanh would be appropriate when sigmoid saturation is stalling learning, not for unstable loss.
B
Add dropout after each layer
Why it fails: Dropout randomly deactivates neurons to reduce overfitting; it does not address oscillating loss caused by an excessive learning rate. Adding it can even slow convergence. Dropout is the right choice when training accuracy far exceeds validation accuracy, indicating overfitting rather than instability.
C
Decrease the learning rate
Lowering the learning rate reduces the size of each gradient descent step, preventing the optimiser from overshooting minima that cause the loss to oscillate rather than converge. For a binary classification network with fluctuating, non-converging loss, this directly addresses the instability constraint in the stem.
D
Increase the number of units in the first dense layer
Why it fails: Adding units increases model capacity, which typically worsens divergence when the learning rate is too high, since larger gradients amplify each update. Capacity changes target underfitting. Increasing layer width is correct when the model cannot fit the training data, not when loss fluctuates without converging.

Full breakdown with real-world context →

Question 7mediummultiple choice

Full question →

Based on the exhibit, what is the most likely cause of the pod failure and its solution?

Exhibit

Refer to the exhibit.

$ kubectl get pods
NAME                     READY   STATUS      RESTARTS   AGE
ml-service-7b9c8f-2k4d   0/1     OOMKilled   3          5m
ml-service-7b9c8f-j5p1   1/1     Running     0          10m

$ kubectl logs ml-service-7b9c8f-2k4d
2025/03/15 14:23:45 [FATAL] Out of memory: Killed process 1234 (python)

A
The node has insufficient CPU; add more CPU.
Why it fails: Insufficient CPU produces pending or throttled pods, not the crash or image-pull failure shown in the exhibit. It is tempting because CPU requests are a common scheduling cause, and would be correct if the pod events showed FailedScheduling with insufficient cpu on the node.
B
The pod is configured with wrong GPU drivers; update drivers.
Why it fails: Wrong GPU drivers cause runtime errors only when the container actually requests a GPU; the exhibit shows a different failure signature. It is tempting because driver mismatches genuinely break GPU workloads, and would be correct if logs showed CUDA initialisation errors after a successful image pull.
C
The model is too large; use a smaller model.
Why it fails: Model size affects memory and load time, not the failure state shown in the exhibit. It is tempting because oversized models do cause OOMKilled pods, and would be correct if the exhibit showed memory-limit exhaustion during model loading rather than the actual reported condition.
D
The container memory limit is too low; increase the memory limit in the pod spec.
The container exceeded its configured memory limit, so the kernel terminated it with an OOMKilled status. Raising the memory limit in the pod spec allows the workload to complete, satisfying the resource constraint the exhibit shows the container breaching.

Full breakdown with real-world context →

Question 8easymultiple choice

Full question →

A team is building a regression model to predict house prices. Which data transformation is most appropriate if the target variable exhibits right skewness?

A
Principal component analysis (PCA)
Why it fails: PCA reduces dimensionality by creating uncorrelated components; it does not correct skewness in the target. A log or square-root transform addresses right skew directly. PCA is tempting because it is a common preprocessing step, but it is intended for feature reduction and multicollinearity, not distribution shaping of the label.
B
Standardization (Z-score)
Why it fails: Standardization rescales features to zero mean and unit variance; it preserves the target's right skew, so the regression assumption remains violated. It is tempting because it is the standard preprocessing step for distance-based or regularised models, where it would correctly be applied to predictor variables.
C
One-hot encoding
Why it fails: One-hot encoding converts categorical variables into binary columns and cannot reduce right skewness in a continuous target. It is tempting because it is the standard transformation for nominal features, which is the correct scenario for encoding categorical predictors in a regression model.
D
Log transformation
Log transformation compresses the long right tail of a positively skewed target, pulling extreme values toward the mean and stabilising variance. This satisfies the right-skewness constraint, making the target closer to normal so linear regression's residual assumptions hold and predictions are less distorted.

Full breakdown with real-world context →

Question 9mediummultiple choice

Full question →

Based on the exhibit, what is the most likely issue with the trained model?

Exhibit

Refer to the exhibit.

Training log from a binary classification neural network:
Epoch 1/10 - loss: 1.2345, accuracy: 0.6543, val_loss: 1.4567, val_accuracy: 0.6123
Epoch 2/10 - loss: 0.9876, accuracy: 0.7123, val_loss: 1.2345, val_accuracy: 0.6543
Epoch 3/10 - loss: 0.6543, accuracy: 0.8123, val_loss: 1.0123, val_accuracy: 0.7123
Epoch 4/10 - loss: 0.4567, accuracy: 0.8765, val_loss: 0.9876, val_accuracy: 0.7345
Epoch 5/10 - loss: 0.3456, accuracy: 0.9123, val_loss: 0.9567, val_accuracy: 0.7567
Epoch 6/10 - loss: 0.2345, accuracy: 0.9456, val_loss: 0.9345, val_accuracy: 0.7789
Epoch 7/10 - loss: 0.1234, accuracy: 0.9678, val_loss: 0.9123, val_accuracy: 0.7890
Epoch 8/10 - loss: 0.0987, accuracy: 0.9789, val_loss: 0.9012, val_accuracy: 0.7912
Epoch 9/10 - loss: 0.0765, accuracy: 0.9876, val_loss: 0.8956, val_accuracy: 0.7900
Epoch 10/10 - loss: 0.0543, accuracy: 0.9932, val_loss: 0.8876, val_accuracy: 0.7890

A
Overfitting because training accuracy is much higher than validation accuracy
Overfitting occurs when a model memorises training data, producing high training accuracy but poor generalisation, evidenced by much lower validation accuracy. This directly satisfies the exhibit's constraint: the widening gap between the two accuracy curves. Regularisation, dropout, or more training data would reduce this variance.
B
Data leakage artificially inflating training accuracy
Why it fails: Data leakage inflates training accuracy while validation accuracy stays low, yet the exhibit shows training accuracy itself is not high. It is tempting because leakage is a common cause of suspiciously strong training results, but here the training metric is weak, so leakage cannot explain the pattern.
C
Vanishing gradients causing no learning
Why it fails: Vanishing gradients would stall training loss near its initial value with negligible change across epochs, whereas the exhibit shows loss decreasing steadily. It is tempting because vanishing gradients also produce poor convergence, but the mechanism differs: no learning at all versus learning that generalises poorly.
D
Underfitting due to insufficient epochs
Why it fails: The exhibit shows training and validation loss both still falling together, which points to further training helping, not too few epochs. Underfitting from insufficient epochs is tempting when both curves plateau high, but here the gap between curves, not their level, is the diagnostic signal.

Full breakdown with real-world context →

Question 10mediummultiple choice

Full question →

Refer to the exhibit. The training log shows loss and accuracy for a binary classification model. What is the most likely issue with this model?

Exhibit

Training log:
Epoch 1/10 - loss: 0.6932 - accuracy: 0.5023 - val_loss: 0.6941 - val_acc: 0.5001
Epoch 2/10 - loss: 0.6810 - accuracy: 0.5432 - val_loss: 0.7123 - val_acc: 0.4987
Epoch 3/10 - loss: 0.6645 - accuracy: 0.5876 - val_loss: 0.7356 - val_acc: 0.4953
...
Epoch 10/10 - loss: 0.6234 - accuracy: 0.6521 - val_loss: 0.8123 - val_acc: 0.4889

A
Overfitting
Overfitting is indicated when training loss keeps falling and training accuracy approaches 100% while validation loss rises and validation accuracy stagnates or degrades. That divergence between training and validation curves is the classic signature, matching the exhibit's logged behaviour.
B
Insufficient epochs
Why it fails: Insufficient epochs would leave both training and validation loss still decreasing together; the exhibit shows training loss continuing to fall while validation loss climbs, indicating overfitting rather than premature stopping. It is tempting because stopping training early is a common cause of a model that has not yet converged.
C
Underfitting
Why it fails: Underfitting would show both training and validation loss remaining high and accuracy plateauing near chance; the exhibit instead shows training loss falling while validation loss rises, which is overfitting. It is tempting because underfitting is the classic diagnosis when a model's accuracy stalls at a poor level.
D
Data leakage
Why it fails: Data leakage would inflate validation accuracy while training loss falls normally, not produce the exhibit's diverging curves. It is tempting because leakage genuinely causes suspiciously high scores, and it would be the answer if training accuracy were near-perfect while validation accuracy stayed near chance.

Full breakdown with real-world context →

Question 11hardmultiple choice

Full question →

The exhibit shows the output of a drift monitoring command for a fraud detection model. The team has an automated pipeline that triggers retraining when the overall average drift score exceeds 0.10. Based on the exhibit, what should the operations team do next?

Network Topology

A
Force retraining on all features to ensure the model adapts to the new data distribution.
Why it fails: Retraining every feature ignores the exhibit's per-feature drift scores; only features breaching their own thresholds warrant action, and the pipeline's overall-average trigger may not have fired. It tempts because full retraining suits confirmed, broad distribution shift across all features, not isolated drift.
B
Manually analyze the drift in 'amount' and 'location' and investigate potential causes.
The exhibit shows per-feature drift exceeding the threshold on 'amount' and 'location' while the overall average stays below 0.10, so the automated retraining trigger never fires. Manual investigation of those two features is required to determine whether the drift is genuine and warrants intervention.
C
No action is needed because the model is performing within acceptable drift limits.
Why it fails: Doing nothing ignores that the pipeline's configured trigger is an overall average drift score above 0.10, so if the exhibit shows that threshold breached, retraining is required. It is tempting because small per-feature drift values can look harmless, and no action would be correct when the average stays at or below 0.10.
D
Initiate the automated retraining pipeline since the average drift exceeds 0.05.
Why it fails: The pipeline triggers only when the overall average drift score exceeds 0.10, so acting on a 0.05 threshold fires retraining prematurely and wastes compute. It is tempting because 0.05 is a common drift alert threshold, and it would be correct if the team's configured trigger were 0.05 rather than 0.10.

Full breakdown with real-world context →

These AI0-001 practice questions are part of Courseiva's free CompTIA certification practice question bank. Courseiva provides original exam-style AI0-001 questions with detailed explanations, topic-based practice, mock exams, readiness tracking, and study analytics.

Refer to the Exhibit Practice Questions

How to approach refer to the exhibit practice questions

Quick answer

Related AI0-001 topic practice pages

Implementing AI Solutions practice questions

AI Governance and Ethics practice questions

AI Concepts and Techniques practice questions

AI Concepts and Foundations practice questions

AI Security practice questions

AI Infrastructure and Technologies practice questions

AI Models and Data Engineering practice questions

Machine Learning and Deep Learning practice questions

AI Security, Ethics and Governance practice questions

AI Implementation and Operations practice questions

CompTIA A+ hardware practice questions

CompTIA A+ mobile devices practice questions

Practice scenarios

Refer to the exhibit. A data engineer runs a validation report on the customers table. The "income" column has 12 null values. Which imputation strategy is most appropriate for this column?

Exhibit

Refer to the exhibit. An AI auditor reviews the fairness configuration. What is the purpose of this policy?

Exhibit

Refer to the exhibit. What is the recall of the model?

Exhibit

Refer to the exhibit. A team deploys a sentiment analysis model with this policy. After one month, the monitoring system triggers an alert for feature drift. Which action should the team take first?

Exhibit

Refer to the exhibit. A machine learning pipeline configuration is shown. During a deployment, the model evaluation passes with accuracy 0.86 and precision 0.79. However, the pipeline proceeds to deploy. What is the most likely reason for this behavior?

Exhibit

Refer to the exhibit. A data engineer is training a binary classification neural network. The loss fluctuates and does not converge. Which hyperparameter adjustment is most likely to stabilize training?

Exhibit

Based on the exhibit, what is the most likely cause of the pod failure and its solution?

Exhibit

A team is building a regression model to predict house prices. Which data transformation is most appropriate if the target variable exhibits right skewness?

Based on the exhibit, what is the most likely issue with the trained model?

Exhibit

Refer to the exhibit. The training log shows loss and accuracy for a binary classification model. What is the most likely issue with this model?

Exhibit

The exhibit shows the output of a drift monitoring command for a fraud detection model. The team has an automated pipeline that triggers retraining when the overall average drift score exceeds 0.10. Based on the exhibit, what should the operations team do next?