Experimentation
Free NVIDIA-Certified Associate: Generative AI Multimodal practice — 6 questions on Experimentation, with explanations. No sign-up. Full 12-question mixed test →
A team fine-tunes a text-to-image diffusion model. On a held-out set of 10,000 real images vs 10,000 generated images, FID drops from 45 (baseline) to 28 (fine-tuned). What does this lower FID score indicate?
FID (Fréchet Inception Distance) measures the distance between the feature-space statistics (mean and covariance of Inception activations) of real and generated image sets. A lower FID means the two distributions are statistically closer, indicating the generated images are more realistic and representative of the real data distribution.
A company needs a chatbot that answers questions grounded in an internal document repository that changes daily. The team cannot afford to retrain the model every day. Which adaptation strategy BEST meets these requirements?
RAG decouples the model's parametric knowledge from the frequently changing document store by retrieving up-to-date documents at inference time and inserting them into the context. This handles daily updates without any retraining, making it the correct adaptation strategy for rapidly changing knowledge sources.
During fine-tuning of a large multimodal transformer with NeMo, the training loss spikes to NaN within the first 100 steps at a peak learning rate of 3e-4. Which change is MOST likely to fix this instability?
Large transformer models are especially sensitive to high learning rates at the very start of training/fine-tuning, when weights (or newly initialized adapter layers) are far from optimal. A warmup phase gradually ramps the LR from a small value up to the peak, preventing the large, destabilizing gradient updates that cause loss to diverge to NaN — this is standard practice in NeMo fine-tuning recipes.
A text-to-image model produces an image that is highly photorealistic and statistically similar to real photographs, but the image does not depict the object described in the input prompt at all. Which evaluation metric would BEST capture this specific failure?
CLIP score measures the cosine similarity between the CLIP embedding of the generated image and the CLIP embedding of the text prompt, directly quantifying text-image semantic alignment. A photorealistic image that ignores the prompt content would score low on CLIP score even though it looks like a real photo.
A research team wants to isolate the specific contribution of adding an audio encoder to a multimodal model's overall performance gain. Which experimental design correctly implements this ablation study?
A valid ablation isolates the effect of a single variable — here, presence vs absence of the audio encoder — by holding every other factor (data, hyperparameters, training steps, seed) constant across both runs. Only then can the observed performance difference be attributed specifically to the audio encoder.
Which practice, when tracking experiments in NVIDIA NeMo, is essential for ensuring a fine-tuning run can be exactly reproduced later by another team member?
Full reproducibility requires capturing every factor that can affect training outcomes: the random seed (controls initialization and data shuffling), the exact hyperparameter configuration file, the specific dataset version used, and the exact container/framework version (since library updates can change numerical behavior). Logging all of these together is the standard NeMo/NGC experiment tracking practice for reproducible research.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$99.99 $34.99 with code FREETEST33 — valid through September 9.
Get my $34.99 deal →