Free NVIDIA Certified Associate: Generative AI LLMs NCA-GENL practice questions

Sample question 1. A chat service runs a large model whose GPUs are far from fully used at low batch sizes, and per-token latency is the problem. The team must NOT change the distribution of text the large model produces. Which technique is designed for this?

1. Quantize the large model to 4-bit weights to speed up every forward pass.
2. Let a small draft model propose tokens that the large model verifies in one pass.
3. Serve the small draft model instead of the large model for most incoming chat requests.
4. Lower the sampling temperature so that fewer candidate tokens are considered.

Sample question 2. Model X (32k vocabulary) reports perplexity 9.1 and Model Y (128k vocabulary) reports perplexity 14.3 on the same text. A teammate concludes X is the better language model. Why is that conclusion unsafe?

1. Higher perplexity always means a better model on identical text
2. Perplexity is per token, and the models split the same text into different tokens
3. Perplexity can only be computed on training data, not held-out text
4. The models were probably run on different GPU types with different numerics

Sample question 3. An LLM-based classifier flags support tickets that mention a safety incident. Reviewers say it misses too many real incidents, and a few extra false alarms are acceptable. How should the team adjust the decision threshold, and which metric should rise?

1. Raise the threshold so that precision rises.
2. Keep the threshold and report accuracy instead.
3. Lower the sampling temperature so the classifier is more confident.
4. Lower the threshold so that recall rises.