NCA-GENL — Flashcards by Topic and Blueprint Area
Review of the ten topics published by NVIDIA, mapped to the five weighted areas of the blueprint. Question → answer format: cover the answer and respond aloud.
Review of the ten topics published by NVIDIA, mapped to the five weighted areas of the blueprint. Question → answer format: cover the answer and respond aloud.
Blueprint verified on August 21, 2026. Topics may change; verify the official NVIDIA page before taking the exam.
Topic 1 — Fundamentals of ML and Neural Networks#
Q: What differentiates supervised from unsupervised learning? A: Supervised uses labeled data (classification, regression); unsupervised discovers structure without labels (clustering, PCA).
Q: With which paradigm are LLMs pre-trained? A: Self-supervised learning: the label comes from the text itself (predicting the next token).
Q: What does backpropagation do? A: Computes the loss gradients with respect to each weight (chain rule) so the optimizer (SGD/Adam) can update them.
Q: How do you detect overfitting? A: The training loss/metric continues to improve while the validation loss/metric worsens or stagnates.
Q: What does self-attention compute? A: softmax(QKᵀ/√d)·V — weights how much each token attends to others using queries, keys, and values.
Q: Why does the transformer need positional encodings? A: Because self-attention is order-invariant; without positions, the sequence would be a "bag of tokens".
Q: Encoder-only vs. decoder-only: example and use case for each. A: Encoder-only: BERT, for understanding/classification. Decoder-only: GPT/Llama, for autoregressive text generation.
Q: What is BPE? A: Byte-Pair Encoding: a subword tokenization that iteratively merges the most frequent symbol pairs; used by GPT and Llama.
Q: Why do GPUs accelerate deep learning? A: Thousands of cores and high bandwidth execute matrix multiplications in parallel; CUDA/cuDNN expose this parallelism to frameworks.
Q: What are Tensor Cores? A: NVIDIA GPU hardware units specialized in mixed-precision matrix multiplication (FP16/BF16/FP8).
Topic 2 — Prompt Engineering#
Q: What is few-shot prompting? A: Include input→output examples in the prompt so the model learns the pattern in-context, without updating weights.
Q: What is chain-of-thought and when does it help? A: Ask for step-by-step reasoning before the answer; improves arithmetic and logical reasoning tasks.
Q: What is self-consistency? A: Generate multiple reasoning chains via sampling and select the majority answer.
Q: What does temperature control? A: Sampling randomness: low → near-deterministic output; high → more diversity (and higher error risk).
Q: Difference between top-k and top-p? A: Top-k samples from the k most probable tokens (fixed number); top-p samples from the smallest set whose cumulative probability reaches p (variable size).
Q: What is the system prompt for? A: Set the role, tone, and constraints that govern the entire conversation, separate from user turns.
Q: What is prompt injection? A: Malicious text (in user input or retrieved documents) that attempts to override system instructions.
Q: What configuration do you use for a factual data extraction task? A: Low temperature (≈0), clear instructions, structured output format, and delimiters for the data.
Q: Prompt engineering vs. fine-tuning: which do you test first and why? A: Prompt engineering: it's immediate and cheap; fine-tuning only if prompting (and RAG) cannot meet the required quality.
Topic 3 — Alignment#
Q: Order the phases: RLHF, pre-training, SFT. A: Pre-training → SFT (instruction tuning) → RLHF.
Q: What is the reward model in RLHF? A: A model trained on human response comparisons that scores outputs; the policy is optimized (PPO) to maximize that score.
Q: What does DPO simplify compared to RLHF? A: Optimizes directly on preferred/rejected response pairs, without a separate reward model or reinforcement learning loop.
Q: What is RLAIF? A: Like RLHF, but preference feedback is generated by another AI model instead of humans.
Q: What is Constitutional AI? A: The model critiques and revises its own outputs against a written list of principles ("constitution"), reducing reliance on harmful human feedback.
Q: What is a hallucination and a typical mitigation? A: Fluent but factually incorrect output; mitigated with RAG/grounding in sources, low temperature, and factual evaluation.
Q: What is NeMo Guardrails? A: NVIDIA's open-source toolkit for defining programmable rails in conversational apps: allowed topics, safety, jailbreak detection, flow control.
Q: What is red teaming? A: Deliberately attacking the model (jailbreaks, sensitive topics) before production to discover and fix security flaws.
Q: Why does RLHF use a KL penalty? A: To prevent the policy from drifting too far from the SFT model and avoid reward hacking / language degradation.
Topic 4 — Data Analysis and Visualization#
Q: What plot do you use to view the distribution of a numerical variable? A: Histogram (or density plot).
Q: What plot shows median, quartiles, and outliers at a glance? A: Box plot.
Q: What plot do you use for the relationship between two numerical variables? A: Scatter plot.
Q: Mean or median with extreme outliers? A: Median: it is robust to outliers; the mean shifts with them.
Q: What does df.groupby("col").mean() do in pandas?
A: Groups rows by values of col and calculates the mean of the other numerical columns per group.
Q: Does correlation imply causation? A: No: there may be confounders or coincidence; causation requires controlled experiments.
Q: Why would you analyze the distribution of sequence lengths in a corpus? A: To choose a max sequence length and truncation/padding strategy without cutting too much content or wasting compute.
Q: What is cuDF? A: A RAPIDS library with a pandas-like API running on GPU, to accelerate analysis of large datasets.
Q: What visualization would you use for a confusion matrix? A: Heatmap.
Topic 5 — Experimentation#
Q: What is perplexity and what does a low value indicate? A: Exponential of the mean cross-entropy; lower = the model assigns higher probability to the actual text (predicts better).
Q: BLEU vs. ROUGE: orientation and typical use case for each? A: BLEU: n-gram precision, translation. ROUGE: n-gram/subsequence recall, summarization.
Q: What does BERTScore add over BLEU/ROUGE? A: Measures semantic similarity using contextual embeddings, not just literal n-gram overlap.
Q: Precision and recall: short definition of each? A: Precision: of what I predicted positive, how much was actually correct. Recall: of what was actually positive, how much I detected.
Q: Why does accuracy mislead with imbalanced classes? A: Always predicting the majority class yields high accuracy without detecting anything; use precision/recall/F1.
Q: What is the validation set for (vs. test)? A: Tune hyperparameters and early stopping; the test set is reserved for final evaluation, used only once.
Q: What is data leakage? A: Test/future information leaks into training (duplicates, benchmark contamination), inflating metrics.
Q: What is LLM-as-a-judge? A: Using a powerful LLM to score another model's outputs with a rubric; scales open-quality evaluation.
Q: Train loss drops, validation loss rises: diagnosis and remedy? A: Overfitting; early stopping, regularization, more data.
Topic 6 — Data Preprocessing and Feature Engineering#
Q: Standardization vs. min-max normalization. A: Standardization: mean 0, std 1 (z-score). Min-max: rescales to [0,1]. Both are fitted only on train.
Q: Why is the scaler fitted only on train? A: Fitting it on test leaks test statistics into training (data leakage).
Q: When to use one-hot vs. ordinal encoding? A: One-hot for unordered categories (nominal); ordinal only when a true order exists between categories.
Q: What is SMOTE? A: An oversampling technique that synthesizes minority class examples by interpolating between neighbors, for imbalanced datasets.
Q: What do truncation and padding do in tokenization? A: Truncation cuts long sequences to the maximum allowed; padding fills short ones to a fixed length; the attention mask ignores padding.
Q: Why deduplicate the pre-training corpus? A: Reduces memorization, improves generalization, and prevents over-representing repeated content.
Q: What is chunking in RAG and why does size matter? A: Splitting documents into indexable fragments; chunks that are too large dilute relevance, and too small ones lose context.
Q: What is NeMo Curator? A: NVIDIA's GPU-accelerated tool for curating corpora at scale: deduplication, quality/language filtering, PII removal.
Q: What is back-translation? A: NLP data augmentation: translate to another language and back, generating paraphrases of the original text.
Topic 7 — Experiment Design#
Q: Why change only one variable per experiment? A: If you change multiple at once, you cannot attribute the improvement/deterioration to any specific one.
Q: Grid search vs. random search: which performs better with a fixed budget and why? A: Random search, because it explores more values of the dimensions that actually matter in high-dimensional spaces.
Q: What does Bayesian optimization offer? A: Models the hyperparameter→metric relationship and selects the next point to test strategically, reducing expensive evaluations.
Q: What is an ablation study? A: Removing a system component (e.g., a RAG reranker) and measuring the impact to understand its true contribution.
Q: What does an A/B test need to be conclusive? A: Random assignment, a priori defined metric, and a sufficient sample size for statistical significance.
Q: What is early stopping with patience? A: Stop training when the validation metric hasn't improved for N evaluations (patience), keeping the best checkpoint.
Q: Name three things to fix/record for reproducibility. A: Random seeds, data/code/config versions, and experiment tracking (MLflow, W&B).
Q: Why start experiments with small models/datasets? A: Iterate cheaply and quickly to validate the idea before spending full-scale GPU-hours.
Topic 8 — Software Development#
Q: Why pin versions in requirements.txt? A: Reproducibility: the same code with different dependencies may behave differently or break.
Q: Why mock the LLM API in unit tests? A: Determinism, speed, and zero cost; the real model is tested in integration tests/evals.
Q: What does Docker bring to an ML project? A: Packages code, dependencies, and user CUDA libraries into a reproducible image that runs identically on any host.
Q: What is NGC? A: NVIDIA GPU Cloud: a catalog of containers, pre-trained models, and GPU-optimized Helm charts.
Q: What is needed for a Docker container to see the GPU? A: NVIDIA drivers on the host + nvidia-container-toolkit (and requesting the GPU at runtime).
Q: Where are API keys stored? A: In environment variables or a secret manager; never hardcoded or committed to the repo.
Q: What does a typical CI pipeline for an LLM project run? A: Lint/format, unit tests, and model/prompt regression evals on every push or PR.
Q: What pattern do you use against LLM API rate limits? A: Retries with exponential backoff (and jitter), plus timeouts and explicit error handling.
Topic 9 — Python Libraries for LLMs#
Q: What is autograd in PyTorch? A: The automatic differentiation engine that records operations on tensors and computes gradients in the backward pass.
Q: What does pipeline() from Hugging Face Transformers do?
A: Ready-to-use inference for a task (text-generation, sentiment...) encapsulating tokenizer + model + post-processing.
Q: What is LoRA? A: Efficient fine-tuning: freezes base weights and trains added low-rank matrices; trains <1% of parameters with quality near full fine-tuning.
Q: What does QLoRA add over LoRA? A: The base model is loaded quantized to 4-bit, further reducing memory required for fine-tuning.
Q: What is sentence-transformers for? A: Generate sentence/paragraph embeddings for semantic search, clustering, and RAG.
Q: What is FAISS? A: A similarity search (ANN) library for dense vectors, used as a vector index in retrieval.
Q: What role does LangChain play relative to Transformers? A: LangChain orchestrates applications (chains, RAG, agents, memory) over served models; Transformers provides the models and their training/inference.
Q: What is NVIDIA NeMo? A: End-to-end framework for building and customizing LLMs: pre-training, SFT, PEFT, RLHF, with integrated multi-GPU parallelism.
Q: What does model.to("cuda") do?
A: Moves the model weights to GPU memory to compute there (input tensors must go to the same device).
Topic 10 — LLM Integration and Deployment#
Q: What is quantization and its trade-off? A: Reducing the numerical precision of weights/activations (FP16→INT8/INT4): less memory and more speed in exchange for a typically small quality loss.
Q: What is the KV cache? A: Caching keys/values of already processed tokens to avoid recomputing them at each step of autoregressive generation; consumes memory proportional to context×batch.
Q: What is in-flight (continuous) batching? A: Adding and removing sequences from the batch dynamically during generation, maximizing GPU utilization over static batching.
Q: What is TensorRT-LLM? A: NVIDIA's library that optimizes LLMs for inference: kernel fusion, quantization, paged KV cache, in-flight batching, and tensor parallelism.
Q: What is Triton Inference Server? A: Open-source multi-framework inference server (TensorRT, PyTorch, ONNX...) with HTTP/gRPC, dynamic batching, multi-model support, and Prometheus metrics.
Q: How do TensorRT-LLM and Triton combine? A: TensorRT-LLM optimizes and runs the LLM; Triton serves it in production using TensorRT-LLM as a backend.
Q: What is NVIDIA NIM? A: Containerized microservices with pre-optimized models and a standard API (OpenAI-compatible) to deploy inference on any GPU-enabled infrastructure.
Q: What does time-to-first-token measure and why does it matter? A: Latency until the first generated token; dominates the perceived experience in streaming applications.
Q: RAG or fine-tuning for the model to use data that changes daily? A: RAG: retrieves current information on each query without retraining; fine-tuning freezes knowledge into weights.
Q: What would you monitor for an LLM in production? A: Latency (and TTFT), throughput, error rate, cost per request, and quality: input drift, continuous evals, and user feedback.