NVIDIA NCA-AMAD exam · Official objectives · Original practice content · Last reviewed October 16, 2026

Free NCA-AMAD Data Processing and Data Flywheel Practice Questions

The Data Processing and Data Flywheel domain is 17% of the NVIDIA NCA-AMAD exam and covers ETL with Spark and RAPIDS, data quality and privacy, evaluation datasets, data flywheels and production quality monitoring. Review each answer and explanation as you practice.

The NCA-AMAD exam has 50-60 multiple-choice questions plus a hands-on lab in 1 hour. It costs $125, results are pass or fail, and the certification is valid for 2 years.

Take the 10-question NCA-AMAD quick-start test · NCA-AMAD cheat sheet · Create a free account

What Data Processing and Data Flywheel covers on the NCA-AMAD exam

This domain is 17% of the NCA-AMAD exam (about 10 questions in a 60-question mock). NVIDIA's objectives include:

How to study Data Processing and Data Flywheel

NCA-AMAD Data Processing and Data Flywheel sample questions with answers

Data Processing and Data Flywheel

Sample question 1

In feature engineering for tabular data, what does target encoding do?

  1. Groups a numeric column into a small number of bins by rounding its values
  2. Replaces each category with a target statistic computed for that category
  3. Turns each category into its own binary indicator column with a 1 or a 0
  4. Rescales every numeric column to have a mean of zero and a variance of one
Show correct answer and explanation

Correct answer: B. Replaces each category with a target statistic computed for that category

B is correct because target encoding is a groupby aggregation where the aggregated column is the target, such as the mean price for each brand, so a category becomes a number the model can use. A is incorrect because that is binning, which reduces the precision of a numeric column rather than encoding a category with the target. C is incorrect because that describes one-hot encoding, which creates one column per category and never looks at the target. D is incorrect because that is standardization, a scaling step for numeric features that uses neither the target nor the categories.

Data Processing and Data Flywheel

Sample question 2

Checking a new dataset for missing values, outliers, correlations and feature ranges before any model is trained is part of which activity?

  1. Exploratory data analysis
  2. Model ensembling
  3. Hyperparameter tuning
  4. Inference optimization
Show correct answer and explanation

Correct answer: A. Exploratory data analysis

A is correct because exploratory data analysis (EDA) examines the data's quality, distributions and relationships before modeling, and these checks are its basic steps. B is incorrect because ensembling combines the predictions of several trained models, such as by hill climbing, so it needs trained models first. C is incorrect because hyperparameter tuning searches model settings after a model and a validation scheme already exist. D is incorrect because inference optimization speeds up a deployed model and does not examine the training data.

Data Processing and Data Flywheel

Sample question 3

A data scientist at a retail chain adds a feature and then cross-validates a weekly sales model:

# sales is the target column
df["store_avg_sales"] = df.groupby("store_id")["sales"].transform("mean")
features = ["store_id", "promo", "store_avg_sales"]
scores = cross_val_score(model, df[features], df["sales"], cv=5)

Validation scores are far better than results on the following month's data. What is the MOST likely cause?

  1. The target mean used all rows, so labels from validation folds leak into it
  2. The transform call should be replaced with agg so that the rows align
  3. A groupby mean adds too little information to change the model's accuracy
  4. Five folds are too few, so the validation estimate has very high variance
Show correct answer and explanation

Correct answer: A. The target mean used all rows, so labels from validation folds leak into it

A is correct because when the aggregated column is the target, the statistic must be computed inside each training fold (nested cross-validation); computing it on all rows lets validation rows see their own sales, which inflates the scores. B is incorrect because transform already returns one value per row aligned to the frame; switching to agg changes the shape, not the leakage. C is incorrect because the feature is highly informative precisely because it contains the target values of the validation rows. D is incorrect because high variance would make scores noisy in both directions, not consistently better than live results.

Data Processing and Data Flywheel

Sample question 4

In an insurance pricing dataset, eight categorical columns have missing values, and EDA shows that the pattern of which columns are missing predicts the claim amount. Which engineered feature captures that pattern BEST in a single column?

  1. Drop every row that has a missing value in any of the eight columns
  2. Fill every missing value with that column's most frequent category
  3. Add one boolean column that is true when any of the columns is missing
  4. Encode which of the eight columns are missing as one integer per row
Show correct answer and explanation

Correct answer: D. Encode which of the eight columns are missing as one integer per row

D is correct because giving each column's missing flag its own bit (for example, the flag times 2 to the power of the column's index) and summing them gives a unique integer for every combination of missing columns, which can then feed groupby features. A is incorrect because dropping rows discards data and removes the pattern from the training set entirely. B is incorrect because imputing with the mode hides the missingness, which is exactly the signal the team wants to keep. C is incorrect because a single any-missing flag loses which columns are missing, so different patterns collapse into one value.

Data Processing and Data Flywheel

Sample question 5

A telecom team's gradient-boosted tree model predicts customer churn from a table of account and usage columns, and its accuracy has plateaued. Which next step is MOST likely to improve accuracy?

  1. Engineer groupby aggregation features and keep those that improve validation
  2. Replace the trees with a large neural network trained on the raw columns
  3. Add thousands of generated features at once without checking validation
  4. Remove the most predictive column to stop the model from overfitting
Show correct answer and explanation

Correct answer: A. Engineer groupby aggregation features and keep those that improve validation

A is correct because on tabular data, gradient-boosted trees still gain a lot from crafted features, and groupby aggregations (mean, count or std of one column per group of another) are the most powerful technique; validating each set keeps only real gains. B is incorrect because neural networks extract patterns from raw inputs well in NLP and vision, but on tabular data, tree models with well-engineered features usually win. C is incorrect because untested features add noise and overfitting risk; the winning approach generates many but keeps only the best validated ones. D is incorrect because dropping the strongest signal lowers accuracy; overfitting is controlled with validation and regularization, not by removing good features.

Data Processing and Data Flywheel

Sample question 6

A payments company computes a "transactions in the last hour" feature with SQL for training and with separate Java code in its live scoring service. The model scores well offline but poorly in production, and the two values often differ for the same customer. Which change BEST fixes the root cause?

  1. Increase the model size so it can tolerate noisier input features
  2. Retrain the model every night on the most recent month of transactions
  3. Drop the feature from both pipelines and retrain without it
  4. Define the feature once in a feature store that serves training and scoring
Show correct answer and explanation

Correct answer: D. Define the feature once in a feature store that serves training and scoring

D is correct because a feature store keeps one definition of each feature and serves the same values to offline training and online inference, which removes the training-serving skew caused by two separate implementations. A is incorrect because a larger model still learns from one definition and is scored with another, so the skew remains. B is incorrect because fresher training data does not fix the mismatch between how the feature is computed offline and online. C is incorrect because it removes the symptom by discarding a useful signal instead of making its computation consistent.

Data Processing and Data Flywheel

Sample question 7

During EDA for a grocery demand forecast, the team plots weekly sales and finds a strong upward trend with seasonal peaks. The model will forecast the next eight weeks. Which validation scheme is MOST appropriate?

  1. Train on earlier weeks and validate on the latest weeks, in time order
  2. A single random 80/20 split that is repeated with several seeds
  3. Stratified k-fold by store so that every fold has each store in it
  4. Random 5-fold cross-validation across all weeks to use every row
Show correct answer and explanation

Correct answer: A. Train on earlier weeks and validate on the latest weeks, in time order

A is correct because with a trend and seasonality in the target, time-aware validation mimics deployment by predicting later periods from earlier ones; random splits let the model see the future and look more accurate than it is. B is incorrect because repeating random splits still ignores time order and gives the same optimistic estimate. C is incorrect because balancing stores across folds does not stop future data leaking into training; the issue is time order. D is incorrect because random folds mix future weeks into training, so the scores overstate how well the model forecasts unseen periods.

Data Processing and Data Flywheel

Sample question 8

A data scientist has a pandas notebook that generates and tests candidate features for a pricing model. Testing all 5,000 planned feature ideas on the full dataset would take weeks on CPU, the deadline is in four days, and the team does not want to rewrite the code. Which approach is BEST?

  1. Test only the 50 ideas the team believes are most promising
  2. Test the ideas on a 1 percent random sample of rows on the CPU
  3. Rewrite the pipeline in PySpark and run it on a CPU cluster
  4. Run the same pandas notebook on a GPU with cudf.pandas enabled
Show correct answer and explanation

Correct answer: D. Run the same pandas notebook on a GPU with cudf.pandas enabled

D is correct because cudf.pandas runs existing pandas code on the GPU with zero code changes, which is how thousands of features can be generated and validated in days instead of months. A is incorrect because cutting the search to a guess discards most ideas; the winning results came from exploring thousands of features. B is incorrect because a tiny sample makes feature validation noisy, so the team may keep weak features and miss good ones. C is incorrect because a rewrite conflicts with the no-rewrite constraint and would consume much of the four days.

Data Processing and Data Flywheel

Sample question 9

A bank's credit model scores well in random cross-validation. EDA then shows that the applications it will score next quarter have incomes concentrated well above the range seen in training, with little overlap. What is the MOST justified conclusion before deployment?

  1. Tree-based models are not affected by shifts in feature distributions
  2. The model is safe to deploy because cross-validation was already used
  3. The incoming applications must have been labeled incorrectly
  4. Validation scores likely overstate accuracy on the incoming applications
Show correct answer and explanation

Correct answer: D. Validation scores likely overstate accuracy on the incoming applications

D is correct because a train versus test distribution shift means the validation data resembles the training data but not the deployment data, so a model can validate well and still fail in deployment. A is incorrect because any model trained on one range of values can perform poorly when inputs move outside that range. B is incorrect because cross-validation only measures performance on data drawn like the training set; it cannot reveal a shift it never sees. C is incorrect because a shift in an input feature says nothing about label quality; it is a change in the population being scored.

Data Processing and Data Flywheel

Sample question 10

What does ML lineage tracking record for a machine learning workflow?

  1. Response latency and error rates of the deployed model endpoints
  2. GPU utilization and memory use for every training and inference job
  3. Which datasets, jobs, parameters and models each step used and produced
  4. Which users are allowed to read each dataset and call each endpoint
Show correct answer and explanation

Correct answer: C. Which datasets, jobs, parameters and models each step used and produced

C is correct because lineage tracking stores metadata for every step from data preparation to deployment, so a team can find the exact artifacts behind a model, reproduce it and support audits. A is incorrect because latency and error rates are operational metrics; they do not show where a model's data came from. B is incorrect because resource metrics come from monitoring tools; lineage links artifacts to the steps that used or created them. D is incorrect because access permissions are managed with access control such as RBAC, not with lineage metadata.

Keep practicing

NCA-AMAD practice test hub · Mixed quick start · NCA-GENL practice test

FlashGenius is an independent study resource and is not affiliated with or endorsed by NVIDIA. NVIDIA and NCA-AMAD are trademarks of NVIDIA Corporation. 600 original practice questions.