NCA-GENL Core Machine Learning and AI Knowledge Practice Test 5 — 15 Questions

Question 1

A data scientist is preprocessing a text corpus to train a large language model. They want to convert each word into a dense vector representation that captures semantic relationships before feeding it into the transformer. Which technique should they use?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Word embeddings (e.g., Word2Vec, GloVe). Word embeddings such as Word2Vec or GloVe convert words into dense vectors that encode semantic relationships, which is essential for language models to understand meaning. Tokenization is only a splitting step, while one-hot and TF-IDF yield sparse vectors lacking semantic depth. Thus, embeddings are the correct choice for capturing semantics before transformer processing.

Answer

TF-IDF vectorization

Answer

Tokenization

Answer

One-hot encoding

Question 2

A machine learning engineer is training a large language model and notices that the model performs exceptionally well on the training data but poorly on a held-out test set. Which technique is most appropriate to mitigate this issue?

Accepted Answer

Apply dropout regularization. Overfitting is characterized by high training performance and poor test performance. Dropout regularization mitigates this by preventing the network from relying on specific neurons, encouraging more robust features. Increasing model size, reducing data, or training longer would all aggravate the problem. Therefore, dropout is the appropriate technique.

Answer

Reduce the size of the training dataset

Answer

Train for more epochs

Answer

Increase the model's parameter count

Question 3

A machine learning engineer is training a large language model on a cluster of NVIDIA GPUs. During training, she observes that the loss occasionally spikes to NaN, causing the training to fail. She suspects that the issue is related to the numerical precision of the computations. Which technique is most appropriate to mitigate this issue while maintaining training stability?

Accepted Answer

Use gradient clipping. Gradient clipping is a widely used technique to prevent exploding gradients, which are a common cause of NaN losses in deep learning, especially when training large models with mixed precision. By capping gradient norms, it ensures that parameter updates remain stable, allowing training to proceed without numerical overflow.

Answer

Switch to double precision

Answer

Increase the learning rate

Answer

Reduce the batch size

Question 4

A data scientist is preparing a labeled dataset of 50,000 customer support tickets for supervised fine-tuning of an LLM. Each ticket must be assigned exactly one of eight department labels. Which loss function is most appropriate for training this classification head?

Accepted Answer

Categorical cross-entropy loss. Because every support ticket carries exactly one of eight mutually exclusive department labels, the task is single-label multi-class classification. Categorical cross-entropy, paired with a softmax output layer, directly maximizes the probability of the correct department while normalizing across all eight classes, giving the strongest and most stable training signal for this scenario.

Answer

Mean squared error loss

Answer

Contrastive loss

Answer

Binary cross-entropy loss

Question 5

A machine learning engineer is training a deep neural network and notices that the training loss decreases but the validation loss starts to increase after several epochs. Which two techniques are most appropriate to mitigate this issue? (Choose two.)

Accepted Answer

Use early stopping based on validation loss.. When validation loss increases while training loss decreases, the model is overfitting. L2 regularization penalizes complexity, and early stopping halts training at the optimal point. Both techniques reduce overfitting and improve generalization. Other options either worsen overfitting or are unrelated to the problem.

Answer

Increase the number of training epochs.

Answer

Increase the learning rate.

Answer

Add more layers to the neural network.

Question 6

A team is deploying a large language model for real-time text generation. They observe that the model sometimes produces repetitive and dull outputs, especially when generating longer sequences. They want to encourage more diverse and creative text without significantly degrading coherence. Which decoding strategy should they consider?

Accepted Answer

Top-k sampling. Top-k sampling introduces controlled randomness by sampling from the top k tokens, which increases diversity and reduces repetition. It avoids the pitfalls of greedy and beam search, which are deterministic and prone to dull outputs, while preventing the incoherence that can come from sampling the entire distribution.

Answer

Temperature scaling with a low temperature

Answer

Beam search

Answer

Greedy search

Question 7

A data science team is building a model to predict whether a customer will churn based on historical account activity. They have a large dataset with labeled outcomes (churned or not churned). Which type of machine learning is most appropriate for this task?

Accepted Answer

Supervised learning with a classification algorithm. The task is to predict a binary outcome (churn or no churn) using historical data where the outcome is known. This is a classic supervised learning problem, and classification algorithms are designed to learn from labeled examples to predict discrete categories. Unsupervised, reinforcement, and semi-supervised methods do not directly leverage the available labels for prediction.

Answer

Unsupervised learning with clustering

Answer

Reinforcement learning with a reward function

Answer

Semi-supervised learning with a small labeled set

Question 8

An ML engineer trains a sentiment classifier on 10,000 movie reviews but only 300 are negative. The model predicts positive for nearly every review, including obvious negative ones. Which technique best addresses this class imbalance during training?

Accepted Answer

Apply class weighting in the loss function to penalize errors on the minority class more heavily. With only 300 negative examples against 9,700 positive ones, an unweighted loss is minimized by predicting the majority class. Applying class weights in the loss raises the cost of minority-class errors, forcing the model to learn features that separate negative reviews. This is the most direct and least invasive fix for the skewed decision boundary.

Answer

Normalize the input text by lowercasing and removing stopwords

Answer

Add L2 regularization to all model weights

Answer

Increase the learning rate and train for more epochs

Question 9

A machine learning engineer is training a deep neural network for image classification. They notice that the training loss decreases steadily, but the validation loss starts to increase after a few epochs. Which technique is most directly aimed at addressing this issue?

Accepted Answer

Adding dropout layers. The described behavior—training loss decreasing while validation loss increases—is a classic sign of overfitting. Dropout is a regularization method that randomly drops units during training, which reduces the network's capacity to memorize noise and encourages more generalizable representations. Other options either do not address overfitting or would exacerbate the problem.

Answer

Increasing the learning rate

Answer

Removing batch normalization

Answer

Reducing the size of the training set

Question 10

A data science team at a retail company is building a neural network to predict customer churn from tabular data with mixed numerical and categorical features. They want the model to output a probability between 0 and 1, and they are training with a standard gradient descent optimizer. Which loss function is most appropriate for this binary classification task?

Accepted Answer

Binary cross-entropy loss. Binary cross-entropy is the standard loss for binary classification with a sigmoid output. It directly optimizes the log-likelihood of the correct class, yielding strong gradients when predictions are wrong and well-calibrated probabilities. The other losses either assume multiclass targets, produce non-probabilistic outputs, or suffer from vanishing gradients with sigmoid units.

Answer

Categorical cross-entropy loss

Answer

Mean squared error loss

Answer

Hinge loss

Question 11

A team is deploying a large language model for real-time text generation and notices that inference latency is too high. They want to reduce latency without retraining the model. Which technique is most appropriate?

Accepted Answer

Apply quantization to reduce the model's precision from FP32 to FP16 or INT8.. Quantization reduces the precision of model parameters, which lowers memory bandwidth and speeds up computation. It can be applied without retraining and is a common technique for reducing inference latency. Other options either increase latency or do not affect it. Thus, quantization is the most appropriate choice.

Answer

Increase the batch size for inference requests.

Answer

Use a larger model with more parameters.

Answer

Train the model for additional epochs to improve its efficiency.

Question 12

A team is pretraining a large language model on a cluster of NVIDIA GPUs. They observe that the model's training loss decreases steadily for the first few epochs but then suddenly spikes and eventually becomes NaN. They suspect this is due to exploding gradients. Which technique is most appropriate to address this issue?

Accepted Answer

Apply gradient clipping to limit the magnitude of gradients during backpropagation.. Exploding gradients cause large parameter updates that can destabilize training, leading to loss spikes and NaN values. Gradient clipping is a standard technique that caps gradient norms or values before the optimizer step, preventing such destructive updates. The other options either exacerbate the problem or do not directly target the root cause, making gradient clipping the most appropriate remedy for this scenario.

Answer

Reduce the batch size to decrease the variance of gradient estimates.

Answer

Increase the learning rate to help the model escape the spiking region.

Answer

Switch from Adam to stochastic gradient descent (SGD) without momentum.

Question 13

A team trains a transformer language model on a large corpus but the model achieves very low training loss while performing poorly on held-out text. Which action most directly addresses this outcome?

Accepted Answer

Apply dropout and weight decay to regularize the model. Low training loss combined with weak held-out performance indicates overfitting, meaning the model memorizes training text instead of learning generalizable patterns. Regularization techniques such as dropout and weight decay constrain model capacity and penalize reliance on specific training examples, which directly reduces the train-validation gap and improves performance on unseen text.

Answer

Lower the learning rate for the final training steps

Answer

Increase the number of training epochs on the same corpus

Answer

Reduce the size of the training corpus

Question 14

A machine learning engineer is training a convolutional neural network for image classification and notices that the training loss decreases steadily, but the validation loss starts increasing after a few epochs. The training set is large and representative. Which technique is most directly aimed at addressing this phenomenon?

Accepted Answer

Add dropout layers. The described behavior—training loss falling while validation loss rises—is classic overfitting. Dropout is a direct regularization technique that prevents co-adaptation of neurons and improves generalization. Increasing learning rate, reducing training data, or changing batch size do not specifically counter overfitting and may worsen the outcome.

Answer

Reduce the size of the training set

Answer

Use a smaller batch size

Answer

Increase the learning rate

Question 15

A team is preparing a dataset to train a generative AI model for text summarization. They want to ensure the model generalizes well and does not simply memorize the training examples. Which TWO practices should they follow? (Choose two.)

Accepted Answer

Split the dataset into training, validation, and test sets. To promote generalization and prevent memorization, the team should use a held-out validation set for tuning and a test set for final evaluation, and they should augment training data to increase diversity. Using all data for training, enlarging the model without regularization, or training until training loss is near zero all increase the risk of overfitting.

Answer

Increase the model's parameter count to capture more complex patterns

Answer

Use the entire dataset for training to maximize data availability

Answer

Train for as many epochs as possible until training loss is near zero