NCA-GENL Practice Test

NCA-GENL practice test: what's covered

• Core ML
• Transformers & LLMs
• Prompt Engineering
• RAG
• Experimentation
• NeMo / Triton / TensorRT-LLM
• Trustworthy AI

Free NCA-GENL sample questions

1. A data scientist trains a model that scores 99% accuracy on the training set but only 62% on a held-out validation set. Which condition best describes this behavior and the appropriate first remedy?

   • A. Overfitting — add regularization (e.g., dropout or weight decay) or more training data
   • B. Underfitting — increase model capacity by adding more layers
   • C. Data leakage — the validation labels were exposed during training
   • D. Vanishing gradients — switch the optimizer to plain SGD
   Show answer

   Correct: A. A large gap with high train and low validation accuracy is the signature of overfitting; regularization and more data reduce variance. Underfitting shows low train accuracy too, and leakage typically inflates validation rather than deflating it.

2. In a feed-forward neural network, what is the primary role of a non-linear activation function such as ReLU or GELU between layers?

   • A. It lets the network approximate non-linear functions instead of collapsing into a single linear transform
   • B. It normalizes the input features to zero mean and unit variance
   • C. It guarantees the loss surface is convex so training always finds the global minimum
   • D. It reduces the parameter count by pruning unused weights
   Show answer

   Correct: A. Without non-linearities, stacking linear layers is equivalent to one linear layer. Normalization is a separate operation, activations do not make the loss convex, and they do not prune parameters.

3. A team uses 5-fold cross-validation to evaluate a classifier on a small dataset. What is the main benefit over a single train/test split?

   • A. Every example is used for both training and validation across folds, yielding a more stable performance estimate
   • B. It eliminates the need for a separate test set entirely
   • C. It always increases final model accuracy on unseen production data
   • D. It removes class imbalance from the dataset automatically
   Show answer

   Correct: A. Cross-validation rotates the validation fold so all data contributes to both roles, lowering estimate variance. It does not raise accuracy, balance classes, or replace a final held-out test set.

4. Which statement correctly distinguishes supervised from self-supervised learning as used to pre-train large language models?

   • A. Self-supervised pre-training creates labels from the data itself (e.g., predicting the next token), while supervised learning requires externally provided labels
   • B. Supervised learning needs no labels, while self-supervised learning requires human-annotated labels
   • C. Both require the same human-labeled corpora; the only difference is batch size
   • D. Self-supervised learning can only be applied to image data, never to text
   Show answer

   Correct: A. LLM pre-training is self-supervised: the next-token target is derived from raw text, needing no manual labels. Supervised learning needs explicit labels, and self-supervision applies across modalities.

5. A regression model must penalize large errors much more heavily than small ones during training. Which loss function best fits this requirement?

   • A. Mean Squared Error (MSE)
   • B. Mean Absolute Error (MAE)
   • C. Hinge loss
   • D. Categorical cross-entropy
   Show answer

   Correct: A. MSE squares residuals so large errors dominate the gradient, penalizing them more than MAE (linear). Hinge loss is for margin classifiers and cross-entropy is for classification, not regression.

Last updated: 2026-10-19

NCA-GENL practice test FAQ