NCA-GENL Mock Questions

Correct: Next-token prediction on unlabeled text

LLM pretraining typically uses next-token prediction built from raw text without manual labels.

Correct: Intent classifier that routes tickets

Classification/routing is a discriminative task; answer generation is generative.

Correct: Lower adaptation time and compute cost

Starting from pretrained weights reduces data, time, and compute needed for downstream tasks.

Correct: Cross-entropy

Cross-entropy aligns naturally with probabilistic next-token prediction.

Correct: Gradient clipping

Gradient clipping limits update magnitude and helps prevent unstable loss spikes.

Correct: Decoder-only

Decoder-only models predict the next token from prior context and are common for chat.

Correct: To add token order information

Attention alone is permutation-invariant, so position signals are needed for sequence order.

Correct: It allows multiple relational views in parallel

Different heads can learn complementary relationships such as syntax and long-range dependencies.

Correct: Long context can still degrade quality if retrieval is noisy

Large windows still require relevant, clean context; irrelevant chunks can hurt answer quality.

Correct: Generation latency

Reusing prior key/value states avoids redundant computation and reduces token generation latency.

Correct: Pretraining -> adaptation (SFT/PEFT) -> deployment

General capability is built in pretraining, then adapted and deployed for specific tasks.

Correct: Labeled input-output examples

SFT uses curated labeled examples to shape task behavior and response style.

Correct: Improve instruction following and response formatting

Instruction tuning improves how reliably the model follows user intent and format constraints.

Correct: Replicates model copies and splits batches across workers

Data parallelism keeps model replicas on workers and distributes input batches.

Correct: When a model does not fit on a single device

Model parallelism partitions large models across multiple devices when single-device memory is insufficient.

Correct: Lower memory use and faster training

FP16/BF16 usually reduce memory pressure and improve throughput.

Correct: Emulates larger effective batch size with limited memory

It accumulates gradients over micro-batches before update steps.

Correct: It enables recovery and reproducibility

Checkpoints allow restart after failure and provide auditable run artifacts.

Correct: Decoupled weight decay from gradient update

AdamW applies weight decay in a decoupled way and is widely used in modern training pipelines.

Correct: To stabilize early optimization before full step sizes

Warmup prevents unstable early updates by ramping learning rate gradually.

Correct: When output format consistency is important

Few-shot examples anchor format and behavior for tasks with strict output expectations.

Correct: High-level behavior and policy instruction layer

System prompts define global instructions and boundaries for model behavior.

Correct: More sampling randomness

Higher temperature increases variability in token sampling.

Correct: Selects from top k probable tokens each step

Top-k restricts sampling to a fixed-size candidate set of likely tokens.

Correct: Potentially better sequence likelihood with higher compute cost

Beam search explores multiple candidate sequences but costs more compute.

Correct: Greedy decoding with fixed prompts

Greedy decoding avoids random sampling and is typically repeatable with fixed conditions.

Correct: Treat retrieved content as untrusted and enforce policy boundaries

Injection defense depends on trust boundaries, sanitization, and strict policy enforcement.

Correct: To improve downstream parsing reliability

Schema-constrained outputs reduce parser failures and integration errors.

Correct: Map model output to explicit tool/action interfaces

Function calling turns model intent into structured, callable actions.

Correct: Grounding responses with external knowledge

RAG connects generation to retrievable evidence for fresher and auditable answers.

Correct: A dense vector representing semantic meaning

Embeddings map text (or other data) into vectors used for similarity search.

Correct: To preserve context across chunk boundaries

Overlap helps avoid losing important context that spans adjacent segments.

Correct: Better recall and precision balance

Combining lexical and semantic retrieval can improve both coverage and relevance.

Correct: Improve relevance ordering among initial retrieved candidates

Reranking refines candidate ordering so better evidence reaches generation.

Correct: Limit retrieval to valid source scope or tenant boundaries

Metadata constraints enforce relevance and governance requirements.

Correct: Lower adaptation cost with smaller trainable parameter sets

PEFT updates a small subset of parameters to reduce training and storage cost.

Correct: Low-rank trainable matrices approximate weight deltas

LoRA learns compact low-rank updates rather than full-matrix modifications.

Correct: Virtual prompt embeddings rather than core model weights

Prompt tuning learns input-side embeddings and keeps most model weights frozen.

Correct: Behavior regressions without strong validation

Merged checkpoints can introduce unexpected quality or safety regressions.

Correct: Quality target, compute budget, and deployment constraints

Method selection is a tradeoff between required quality and operational cost/complexity.

Correct: Preference-aligned quality scores for candidate responses

The reward model estimates preference quality and guides policy optimization.

Correct: Train reward model from preference data

Preference data is used to train reward scoring before updating policy behavior.

Correct: Reward hacking

Models may exploit reward-model shortcuts rather than improving true usefulness.

Correct: Principle-based self-critique and revision

Constitutional AI uses explicit principles to critique and refine responses.

Correct: Reduce unsafe or policy-violating outputs

Safety alignment steers behavior toward safer, policy-consistent outputs.

Correct: How well a model predicts token sequences

Perplexity reflects predictive uncertainty over token sequences.

Correct: ROUGE

ROUGE compares overlap patterns and is a common summarization baseline metric.

Correct: Generation throughput speed

Tokens/sec measures how fast the system generates output tokens.

Correct: Higher batching can increase throughput but add latency

Batching improves utilization/throughput but queueing can increase response delay.

Correct: Find plausible outputs unsupported by evidence

Hallucination checks target unsupported claims that sound confident but are not grounded.

Correct: Stability under noisy or shifted inputs

Robustness evaluates how performance holds under input variation and distribution shift.

Correct: It intentionally uses difficult or malicious inputs

Adversarial tests intentionally stress weak points in behavior and controls.

Correct: Dataset auditing and balanced sampling

Data quality and representation checks are foundational bias mitigation controls.

Correct: Both input and output filtering stages

Applying toxicity checks at ingress and egress reduces harmful content risk.

Correct: Policy and rule layers constraining model and tool behavior

Guardrails enforce operational and policy boundaries around model actions.

Correct: An adversarial prompt trying to bypass safety policy

Jailbreak prompts attempt to circumvent refusal and safety constraints.

Correct: Access controls plus data minimization and redaction

Privacy protection depends on strict data handling controls, not decoding settings.

Correct: Ownership, approvals, audit trails, and change control

Governance establishes accountability and traceability across model lifecycle decisions.

Correct: Requirements vary by region and industry

Regulatory obligations differ by jurisdiction, sector, and data sensitivity.

Correct: Continuous correction and iterative behavior improvement

Feedback loops close the gap between observed production behavior and target behavior.

Correct: A model that can process or generate across multiple data modalities

Multimodal systems handle combinations of text, image, audio, or video.

Correct: Answering questions about an image

VLMs jointly reason across visual and textual inputs.

Correct: Iterative denoising from noise to structured output

Diffusion models repeatedly denoise latent noise into coherent outputs.

Correct: Generator versus discriminator competition

GANs use adversarial training between a generator and discriminator.

Correct: They enable retrieval across text and media in a shared space

Shared embedding spaces connect semantics across different modalities.

Correct: Generate descriptive text from visual input

Image captioning maps visual content into textual descriptions.

Correct: Lower precision can improve speed and reduce memory use

Quantization reduces numeric precision to improve efficiency, with quality tradeoffs to validate.

Correct: Remove less important parameters to reduce serving cost

Pruning targets redundant parameters to improve efficiency.

Correct: A smaller student model learns behavior from a larger teacher

Distillation transfers useful behavior into a smaller, cheaper model.

Correct: Optimizing inference execution on NVIDIA hardware

TensorRT compiles and optimizes inference graphs/kernels for GPU performance.

Correct: Streaming inference with incremental token output

Streaming returns partial output quickly, improving user-perceived responsiveness.

Correct: Dynamic capacity adjustment based on load

Autoscaling increases or decreases serving resources as demand changes.

Correct: Serving models in production with backend/runtime integration

Triton is a production serving platform supporting multiple model backends.

Correct: Packaged inference microservices for simpler deployment

NIM packages model serving components to speed practical deployment.

Correct: Out-of-memory failures and fragmentation-related instability

Managing allocation and fragmentation is key for stable high-throughput serving.

Correct: Transform

Extract pulls data, transform reshapes/cleans it, and load stores processed outputs.

Correct: It enables reproducibility and auditability

Versioned datasets let teams reproduce results and trace changes over time.

Correct: Parameters, metrics, artifacts, and run context

Complete run context is required for repeatability and debugging.

Correct: Shifts in data distributions or model behavior

Drift checks detect changing conditions that can silently degrade performance.

Correct: Automate validation, packaging, and release of model changes

CI/CD for ML standardizes safe, repeatable model release workflows.