# Free NCA-ADS practice questions

> 10 free NCA-ADS practice questions with the correct answer and an explanation.

Page: https://devopscodelab.com/exams/nvidia-nca-ads/free-practice-questions/
Practice exams: https://devopscodelab.com/exams/nvidia-nca-ads/

### Question 1 of 10

A data scientist has an existing pandas notebook that loads a 20 GB CSV, filters rows and runs several groupby aggregations. They want GPU acceleration with no code changes and automatic CPU fallback for any operation the GPU library does not support. What should they do?

- A. Install the RAPIDS Accelerator for Apache Spark and point the notebook kernel at it
- B. Replace every pandas import with import cudf as pd and rewrite unsupported calls by hand
- C. Load the cudf.pandas extension with %load_ext cudf.pandas before importing pandas
- D. Convert the notebook to Dask DataFrame and call compute() after each aggregation step

**Answer:** C

**Explanation:** The cudf.pandas accelerator mode intercepts pandas calls, runs them on the GPU through cuDF and transparently falls back to CPU pandas when an operation is unsupported, so no code changes are needed. Swapping the import to cuDF works for many calls but gives no automatic fallback, so unsupported methods fail. Dask DataFrame parallelises across CPU partitions and needs code changes. The Spark RAPIDS Accelerator applies to Spark SQL and DataFrame jobs, not to a pandas notebook.

### Question 2 of 10

A model needs a feature from a high-cardinality 'merchant_id' column with 2 million distinct values. One-hot encoding would explode memory. Which encoding is MOST appropriate for a tree-based GPU model while avoiding target leakage?

- A. Target encoding computed once using all labels in the full training set
- B. One-hot encoding followed by PCA down to fifty components
- C. Label encoding sorted by the alphabetical merchant name
- D. Out-of-fold target encoding computed with cross-validation folds

**Answer:** D

**Explanation:** Out-of-fold target encoding replaces each category with a smoothed target statistic learned only from other folds, so it stays compact and avoids leaking each row's own label. One-hot encoding 2 million values is exactly the memory problem being avoided, even if reduced later. Target encoding on the full set leaks the label into the feature and inflates validation scores. Alphabetical label codes are compact but impose an arbitrary order that carries little signal.

### Question 3 of 10

A binary classifier for a rare disease reports 99% accuracy on a test set with 1% positives. Which metric should the team inspect to judge whether the model actually finds positive cases?

- A. Specificity on the negative class
- B. Mean squared error of the class labels
- C. Recall on the positive class
- D. Overall accuracy on a larger test set

**Answer:** C

**Explanation:** Recall measures the share of actual positives the model detects, which is what matters when positives are rare; a model predicting all negatives would score 99% accuracy but 0% recall. A larger test set does not fix accuracy's insensitivity to imbalance. Mean squared error is a regression metric. Specificity measures performance on the negative majority, which is already high.

### Question 4 of 10

Cross-validation on a GPU XGBoost model gives fold AUCs of 0.91, 0.62, 0.90, 0.89 and 0.60. Rows from the same customer appear in several folds. What is the MOST likely issue and fix?

- A. The learning rate is too low; raise it so all folds converge evenly
- B. Customer-level leakage and grouping; use GroupKFold on customer ID
- C. GPU nondeterminism; train on the CPU so every fold gives one score
- D. There are too few folds; switch to 20 folds so the scores average out

**Answer:** B

**Explanation:** Rows from the same customer in train and validation let the model memorise customers, and uneven fold composition produces unstable scores; GroupKFold keeps each customer in one fold. Learning rate does not explain fold-to-fold swings of this size. More folds hide the variance without fixing leakage. GPU nondeterminism causes tiny differences, not 0.3 AUC gaps.

### Question 5 of 10 (choose 2)

A nightly feature pipeline in Dask-cuDF reads raw Parquet, cleans it, joins three tables and writes features. Runs occasionally produce different row counts from the same input. Which practices make the pipeline reproducible? (Choose two.)

- A. Run the pipeline interactively so engineers can inspect each output by eye
- B. Call persist() on every intermediate DataFrame to keep results in memory
- C. Set explicit random seeds and avoid nondeterministic sampling in the pipeline
- D. Increase the number of Dask partitions so each task processes fewer rows
- E. Pin library versions in an environment file and record it with each run

**Answer:** C, E

**Explanation:** Pinned, recorded dependency versions and fixed seeds remove two common sources of run-to-run variation: changed library behaviour and random sampling. persist() caches data for speed but does not change results. Partition count affects parallelism, not determinism of logic. Manual interactive runs make reproducibility worse, not better.

### Question 6 of 10

An analyst runs a regression and finds a statistically significant coefficient for 'number of support tickets' on churn. A colleague claims tickets cause churn. What is the BEST response?

- A. Agree, because adding more rows to the regression would always confirm the causal effect
- B. Disagree, because regression coefficients cannot be statistically significant when outcomes are binary
- C. Agree, because a statistically significant coefficient proves that the predictor is causal
- D. Association is shown; causation needs an experiment or a causal design accounting for confounders

**Answer:** D

**Explanation:** A significant coefficient shows association conditional on the model, but confounders such as product problems could drive both tickets and churn; causal claims need experiments or causal methods. Significance does not prove causation. Logistic regression coefficients can be significant for binary outcomes. More data narrows intervals but does not establish causality.

### Question 7 of 10

A workflow repeatedly converts a cuDF DataFrame to pandas for one helper function and back to cuDF, and profiling shows most time is spent outside computation. What is the MOST likely bottleneck?

- A. Insufficient CPU cores available for the pandas helper function
- B. Data transfers between host and device memory over the PCIe bus
- C. Disk I/O caused by reading the source Parquet file repeatedly
- D. Kernel launch overhead from too many small GPU threads in each block

**Answer:** B

**Explanation:** Each to_pandas and from_pandas call copies the whole DataFrame between GPU and host memory, and these transfers often cost more than the computation itself. Kernel launch overhead is minor compared with moving gigabytes. CPU core count would show up as computation time in the helper, not transfer time. The scenario describes in-memory round trips, not repeated file reads.

### Question 8 of 10

A deployed churn model's input feature 'monthly_spend' now has a noticeably different distribution than in training, though labels are not yet available. Which check detects this?

- A. Measure average inference latency and alert when it increases sharply
- B. Recompute training accuracy on the original training set every day and alert on any change
- C. Check that the model file checksum still matches the registered version
- D. Compare current and training feature distributions with a statistic such as PSI or a KS test

**Answer:** D

**Explanation:** Data drift is detected by comparing feature distributions between training and production, for example with the population stability index or a Kolmogorov-Smirnov test, and it needs no labels. Training accuracy on old data does not change. A checksum confirms the model file is unchanged, not that inputs are. Latency monitors performance, not input distribution.

### Question 9 of 10

A forecasting team evaluates a model with a rolling-origin approach on GPU. Which description is correct?

- A. Repeatedly train on data up to a cutoff, forecast the next window, then move the cutoff forward
- B. Forecast all windows at once and compare them with the training data used to fit the model
- C. Randomly choose cutoffs, train on data after each cutoff and forecast the earlier windows back
- D. Train once on the full history and evaluate on randomly selected days drawn from that same history

**Answer:** A

**Explanation:** Rolling-origin (walk-forward) evaluation trains on the past up to each cutoff and tests on the following window, then advances, mimicking real forecasting. Training on later data to predict earlier windows reverses time. Evaluating on random days from the training history leaks information. Comparing forecasts with training data measures fit, not forecast skill.

### Question 10 of 10

Which file in a repository tells git to exclude large data files and local environment folders from version control?

- A. .gitattributes
- B. environment.yml
- C. .gitignore
- D. .git/config

**Answer:** C

**Explanation:** .gitignore lists path patterns that git should not track, such as data folders and local environments. .gitattributes sets per-path attributes like line endings or LFS handling. .git/config holds repository settings such as remotes. environment.yml defines a conda environment.

Ready for a timed run? Take the free 15-question NCA-ADS mini exam: https://devopscodelab.com/app/exams/nvidia-nca-ads/mini/

DevOpsCode Lab is not affiliated with NVIDIA. NCA-ADS and related names are trademarks of their owners. We write every question ourselves; these are not real exam questions.
