Fix Prediction Code Quiz Calculator

Advanced Prediction Code Debugging Quiz

Inspect each snippet carefully. Choose the strongest production-safe correction.

Candidate name

Passing target

Practice timer

Apply 25% negative marking

Show detailed answer review

0 of 18 answered

Time: 20:00

Question 1 3 points

Preprocessing

Why does this prediction code produce unstable results after every request?

scaler = StandardScaler()
X_live = scaler.fit_transform(request_rows)
prediction = model.predict(X_live)

The scaler should use fit_predict instead.

The scaler must be fitted again for every user.

Use the scaler fitted during training, then call transform.

Convert the scaled values into strings before prediction.

Question 2 2 points

Shapes

A single sample causes a shape error. Which repair is correct?

sample = np.array([42, 7.5, 1])
prediction = model.predict(sample)

Use sample.reshape(1, -1) before prediction.

Use sample.flatten() before prediction.

Remove the final feature from the sample.

Call model.fit(sample) before prediction.

Question 3 3 points

Probability

The application needs the positive-class probability. What should replace the final line?

proba = classifier.predict_proba(X_live)
risk = proba[0]

risk = proba[:, 0:2]

risk = proba[0, 1]

risk = np.argmax(proba)

risk = classifier.predict(X_live)[1]

Question 4 3 points

Data Integrity

The model receives correct values but wrong predictions. Which fix protects feature order?

payload = {'income': 90000, 'age': 38, 'debt': 12000}
X_live = np.array(list(payload.values()))
prediction = model.predict(X_live.reshape(1, -1))

Sort the values alphabetically without checking training order.

Build a DataFrame using the saved training feature list.

Reverse the values before every prediction.

Convert the dictionary into a set.

Question 5 3 points

Preprocessing

An unseen category crashes the encoder. Which production fix is strongest?

encoder = OneHotEncoder()
X_train = encoder.fit_transform(train[['city']])
X_live = encoder.transform(live[['city']])

Configure handle_unknown="ignore" during training.

Fit the encoder on the live row.

Delete every row containing a new city.

Replace the encoder with random integers.

Question 6 3 points

Framework Mode

Dropout remains active during PyTorch predictions. Which repair is required?

model = torch.load('model.pt')
output = model(batch)

Call model.train() before every prediction.

Call model.eval() and use torch.no_grad().

Increase dropout during inference.

Run optimizer.step() after prediction.

Question 7 2 points

Preprocessing

The target was trained with a logarithmic transform. Which output repair is correct?

y_train_log = np.log1p(y_train)
model.fit(X_train, y_train_log)
y_pred = model.predict(X_live)

Use np.expm1(y_pred) to restore the original scale.

Use np.log1p(y_pred) again.

Round predictions before restoring scale.

Multiply predictions by the feature count.

Question 8 3 points

Probability

A fraud model uses unequal error costs. What is wrong with this decision code?

probability = model.predict_proba(X_live)[:, 1]
flag = probability >= 0.50

Nothing is wrong because 0.50 is always optimal.

The threshold should be selected using validation costs and constraints.

The code should always use the negative-class probability.

Probabilities must be converted into percentages first.

Question 9 3 points

Shapes

Which correction returns one class per sequence from logits shaped batch, time, classes?

logits = model(tokens)
predicted = torch.argmax(logits, dim=1)

Use torch.argmax(logits[:, -1, :], dim=-1).

Use torch.argmax(logits, dim=0).

Flatten logits and choose one global maximum.

Average class indices before returning them.

Question 10 2 points

Data Integrity

Why can this batch code attach predictions to incorrect customers?

clean = customers.dropna().sort_values('income')
preds = model.predict(clean[features])
customers['prediction'] = preds

Predictions are always sorted automatically.

Filtering changes row counts and sorting changes alignment.

The model requires customer names as features.

The prediction column must be created before inference.

Question 11 3 points

Deployment

The service loads a newer model but an older scaler. Which safeguard prevents this?

model = joblib.load('model_v8.pkl')
scaler = joblib.load('scaler_v6.pkl')

Rename both files after every request.

Package version-matched artifacts with metadata and checksums.

Use the oldest model available.

Fit the scaler from one live request.

Question 12 2 points

Data Integrity

A web form submits numeric values as strings. Which repair is safest?

age = request.form['age']
income = request.form['income']
X_live = [[age, income]]

Trust automatic conversion inside every model.

Validate ranges and explicitly convert accepted numeric formats.

Concatenate both strings into one feature.

Use string length as the prediction value.

Question 13 3 points

Preprocessing

Why is this imputation code a leakage risk during evaluation?

combined = pd.concat([train, test])
combined['income'] = combined['income'].fillna(combined['income'].median())

The median calculation uses information from test data.

Median imputation always creates negative values.

Concatenation removes all missing values automatically.

The code should use the test median for training.

Question 14 2 points

Framework Mode

A Keras binary model returns values outside zero and one. What should be checked first?

model.add(Dense(1))
score = model.predict(X_live)

Confirm whether the output is a logit and apply sigmoid when appropriate.

Apply argmax across the single output column.

Divide every output by the batch size.

Replace all negative outputs with class one.

Question 15 3 points

Deployment

A transformer prediction changes when padding length changes. Which input is probably missing?

ids = tokenizer(texts, padding=True, return_tensors='pt')['input_ids']
logits = model(input_ids=ids).logits

The attention mask returned by the tokenizer.

A randomly generated label vector.

A second optimizer instance.

A new vocabulary fitted on each request.

Question 16 2 points

Probability

This multiclass code returns a single class for the entire batch. Which fix is correct?

proba = model.predict_proba(X_batch)
labels = np.argmax(proba)

Use np.argmax(proba, axis=1).

Use np.argmax(proba, axis=0).

Use np.min(proba).

Use labels = proba.shape.

Question 17 3 points

Shapes

A multi-output regressor returns two targets. Why is this line incorrect?

prediction = model.predict(X_live)
price = float(prediction)

The result contains multiple values and needs explicit indexing.

Every regression output must be rounded first.

float only works with classification labels.

The model should be retrained after each request.

Question 18 3 points

Deployment

Which improvement best detects silent prediction failures after release?

prediction = model.predict(X_live)
return {'prediction': prediction.tolist()}

Remove all logs to improve latency.

Monitor input drift, confidence, latency, errors, and delayed outcomes.

Retrain the model after every prediction.

Return only HTTP success responses.

Formula Used

Correct points = sum of weights for correct answers.

Penalty = wrong-answer points × 0.25, when enabled.

Adjusted points = maximum of zero or correct points minus penalty.

Percentage = adjusted points ÷ maximum points × 100.

Unanswered questions receive zero points and no penalty. Category percentages use earned category points divided by available category points.

How to Use This Calculator

Choose a passing target and optional negative marking.
Start the timer when timed practice is useful.
Inspect every code snippet and select one correction.
Enable detailed review when explanations are required.
Press the result button after completing the questions.
Review category scores and repeat weak debugging areas.

Building Reliable Prediction Code

Why Prediction Code Breaks

Prediction code often fails after a model trains successfully. Training success does not guarantee correct production behavior. Small mismatches can silently distort every returned prediction.

Input shape errors are common in deployed applications. Models usually expect rows with fixed feature counts. Extra dimensions may produce failures or misleading outputs.

Feature order also matters during prediction. A correct array can still contain misplaced values. Named columns reduce this risk when pipelines support them.

Preserving Training Logic

Preprocessing must match the training workflow exactly. Scaling values twice changes their learned numerical meaning. Skipping scaling can push inputs outside familiar ranges.

Encoders require the same fitted category mapping. Refitting an encoder creates different numeric representations. Unknown categories need planned handling before deployment begins.

Data leakage can appear inside prediction code. Future information may accidentally enter feature construction. Leakage creates impressive tests but unreliable real outcomes.

Handling Outputs Correctly

Classification models may return probabilities instead of labels. Developers must choose thresholds for operational decisions. Default thresholds rarely fit every business cost.

Multiclass outputs need careful axis handling. The highest probability usually identifies the predicted class. Using the wrong axis returns incorrect labels silently.

Regression predictions require sensible unit checks. A model may output transformed target values. Inverse transformations restore the original business scale.

Protecting Data Quality

Missing values need consistent imputation logic. Production nulls may differ from training nulls. Pipelines should preserve every fitted imputation rule.

Model loading errors can select outdated artifacts. Versioned files help connect code with training data. Metadata should record features, metrics, and dependencies.

Type conversion bugs frequently affect web forms. Numeric strings may contain spaces or separators. Validation should reject malformed inputs before model execution.

Testing Real Workflows

Batch prediction code must preserve row alignment. Sorting or filtering can disconnect outputs from records. Stable identifiers help verify every result assignment.

Prediction services also need range validation. Impossible ages or negative quantities signal bad inputs. Guardrails protect models from unsupported operating conditions.

Reliable tests cover normal and unusual inputs. Include empty values, unseen categories, and extreme ranges. Compare outputs against trusted baseline examples.

Maintaining Production Reliability

Monitoring continues after prediction code ships under real traffic. Data drift can reduce accuracy without exceptions. Track distributions, confidence, latency, and business outcomes.

A debugging quiz builds pattern recognition quickly. Realistic snippets reveal mistakes hidden by clean theory. Immediate explanations connect each fix with its consequence.

Careful prediction code protects users and decisions. Strong pipelines reduce silent failures across changing environments. Regular reviews keep fixes aligned with model updates.

Frequently Asked Questions

What skills does this quiz measure?

It measures debugging across preprocessing, shapes, probabilities, framework modes, data integrity, and deployment. Questions reward production-safe fixes rather than temporary patches.

Why are questions weighted differently?

Some bugs require deeper reasoning or create greater production risk. Weighted scoring gives those scenarios more influence while keeping simpler checks valuable.

How does negative marking work?

When enabled, wrong-answer points create a twenty-five percent penalty. Unanswered questions receive no penalty. The adjusted score never falls below zero.

Can I change the passing score?

Yes. Select sixty, seventy, eighty, or ninety percent before submission. The chosen target determines the displayed pass result.

Does the timer submit automatically?

No. The timer supports practice pacing without forcing submission. Your selected answers remain available until you submit or reset the form.

Why should preprocessing objects be reused?

Fitted preprocessors contain learned statistics, mappings, and feature structures. Refitting them on live data changes the model input meaning and causes inconsistent predictions.

What does category performance show?

It compares earned weighted points with available points inside each debugging category. Low percentages reveal focused areas for further practice.

Are unanswered questions treated as wrong?

They receive zero points but remain separate from wrong answers. Negative marking never applies to unanswered questions.

Can the result be printed?

Yes. Submit the quiz, then use the print result button. The print layout removes interactive controls and preserves key scoring details.

Advanced Prediction Code Debugging Quiz

Formula Used

How to Use This Calculator

Building Reliable Prediction Code

Why Prediction Code Breaks

Preserving Training Logic

Handling Outputs Correctly

Protecting Data Quality

Testing Real Workflows

Maintaining Production Reliability

Frequently Asked Questions

What skills does this quiz measure?

Why are questions weighted differently?

How does negative marking work?

Can I change the passing score?

Does the timer submit automatically?

Why should preprocessing objects be reused?

What does category performance show?

Are unanswered questions treated as wrong?

Can the result be printed?

Related Calculators