Free NCA-AMAD practice test: 30 questions with answers

Predicting a numeric value from labeled examples is supervised regression. Clustering works without labels, reinforcement learning uses rewards, and anomaly detection looks for unusual points.

Overfitting means the model has learned the training data too closely and fails to generalize. Underfitting is the opposite, while normalization and tokenization are preprocessing steps.

The test set gives an unbiased estimate of performance on unseen data. It does not speed training, store weights or replace validation.

Attention computes weighted relationships between tokens so each representation can use context. Compression to one number, punctuation removal and database storage are not its purpose.

A causal mask prevents attention to future positions so the model predicts the next token from the past. Bidirectional masks would reveal future tokens, random order breaks language structure, and vocabulary size is unrelated.

Self-attention is order-agnostic, so positional encodings or embeddings are needed to represent sequence order. Embeddings do not contain order by themselves, and the other options are incorrect.

Periodic evaluation on fixed tests and user feedback reveal quality changes. Restart counts, rack colors and name lengths do not measure quality.

Time to first token measures perceived responsiveness in streaming applications. The other items do not measure that wait.

Canary releases limit exposure while validating the new version on real traffic. Immediate full replacement, deleting first and laptop-only testing increase risk.

Few-shot prompting shows examples in the prompt to guide the output format and behavior. Zero-shot provides no examples, and clipping and distillation are training techniques.

Chain-of-thought prompting elicits step-by-step reasoning, which often improves multi-step accuracy. Random sampling, removing the question and one-word constraints do not elicit reasoning.

An explicit grounding instruction with a fallback for missing answers constrains the model. Creative requests, general-knowledge-first instructions and missing context weaken the constraint.

A vector database stores embeddings and retrieves similar chunks for the query. Training and final generation are performed by the model, and passwords are not stored there.

A grounding check verifies claims against retrieved evidence and reduces unsupported citations. Higher temperature increases randomness, removing citations hides the issue, and images cannot be searched as text.

Chunking keeps passages within model and context limits and improves retrieval precision. Storage limits in general, printing and metadata are not the reason.

A backend service keeps secrets server-side, away from users. Browser code, public forums and page titles expose the key.

Input and output validation with rate limits and authentication reduce misuse. Administrator access, public prompts and disabled logs increase risk.

A clear schema and argument validation make tool calls reliable and safe. Unreviewed code execution, admin credentials and exposed passwords are unsafe.

NIM endpoints provide optimized model access through standard APIs, which speeds development. Drivers, chips and printed weights are not practical development paths.

cuDNN provides optimized primitives for deep learning on GPUs. cuDF handles dataframes, cuGraph handles graph analytics and Omniverse Kit targets 3D application development.

NeMo provides building blocks and recipes for training, customizing and aligning language models. Holoscan, Isaac Sim and Metropolis target other domains.

Deduplication and rebalancing improve training quality and fairness. Training on duplicates, ignoring imbalance and removing labels make results worse.

Detecting and removing or masking personal information protects privacy and compliance before training. Labeling, retaining and printing do not protect the data.

Splitting first prevents information from the test set leaking into training. Computing statistics on all data, copying test examples and choosing splits by score all inflate or distort results.

Whether prompting suffices and whether labeled examples exist drive the choice between approaches. Region names, team sizes and logo colors are irrelevant.

Reviewing and filtering examples improves accuracy and safety of the tuned model. Unreviewed scraping, duplication and removing review degrade quality.

The same held-out set, consistent metrics and human spot checks make comparisons fair. Cherry-picked outputs, different datasets and unevaluated choices are unreliable.

A benchmark that does not match real use, or that overlaps training data, produces misleading scores. Difficulty, layer count and question color do not explain the gap.

Group-level evaluation and human accountability address fairness and responsibility. Assumed objectivity, hidden decisions and averages that hide group differences are problematic.

Informing users and documenting limitations supports informed use. Hiding use, claiming perfection and refusing documentation reduce transparency.