Data Preparation and Fine-Tuning
Free NVIDIA-Certified Professional: Generative AI and LLMs practice — 6 questions on Data Preparation and Fine-Tuning, with explanations. No sign-up. Full 12-question mixed test →
A team is curating a 2TB web-crawled corpus for continued pretraining and needs to remove near-duplicate documents at scale without relying on exact string matching. Which approach is most appropriate?
MinHash+LSH approximates Jaccard similarity between document n-gram sets efficiently at scale, allowing near-duplicate (not just exact) documents to be clustered and filtered — the standard fuzzy deduplication approach used in large-scale corpus curation (e.g., NeMo Curator's fuzzy dedup module).
A company has only 500 labeled instruction-response pairs for a niche legal domain and a single A100 80GB GPU. They want to adapt a 13B-parameter LLM without overfitting or requiring massive compute. Which customization strategy is BEST?
With a tiny labeled dataset and limited compute, LoRA/QLoRA updates only a small number of low-rank adapter parameters, drastically reducing overfitting risk and memory/compute cost compared to full fine-tuning, while still adapting behavior to the instruction format.
During LoRA fine-tuning in NeMo, increasing the rank (r) hyperparameter primarily affects which trade-off?
LoRA rank sets the dimensionality of the low-rank update matrices (A and B); a higher rank means more trainable parameters and greater capacity to fit complex adaptations, but also more memory usage and greater overfitting risk when training data is limited.
A team wants to align a fine-tuned LLM to human preferences but lacks the infrastructure to train and maintain a separate reward model or run PPO-based reinforcement learning. They do have a dataset of paired chosen/rejected responses. Which approach best fits their constraints?
DPO reformulates preference alignment as a classification-style loss directly on chosen/rejected pairs, avoiding the need for a separate reward model and the instability/complexity of PPO-based RL, making it well-suited to teams without RLHF infrastructure.
How does p-tuning differ from LoRA as a parameter-efficient fine-tuning method?
P-tuning trains a small set of continuous virtual token embeddings (via a prompt encoder such as an LSTM/MLP) that are prepended to the input sequence, leaving the base model frozen, whereas LoRA freezes the base weights but injects trainable low-rank matrices into the weight layers (e.g., attention projections) themselves.
When using NVIDIA NeMo Curator to prepare a pretraining dataset, which stage is responsible for identifying and filtering out documents containing toxic or unsafe content before deduplication and training?
NeMo Curator's quality filtering stage applies heuristic rules and trained classifiers (including toxicity/safety and quality classifiers) to score and remove undesirable documents from the corpus before it moves on to deduplication and downstream blending steps.
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.
$129.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →