Core Machine Learning and AI Knowledge

NCA-GENL · 90 questions

• A recycling depot already scores bin weight with a small fitted predictor. A vendor then offers a many-layer network on camera frames of the sorting line, and a third pitch offers a text model that drafts weekly recycling memos. How are classical machine learning, deep learning, and a generative LLM related?
• A dock-permit booth has six numeric columns and a few hundred labeled sailings. A consultant insists they must train a large neural network “because AI.” What should they do first?
• An oyster-farm intern draws a single unit that multiplies incoming measurements by learned numbers, adds a bias, and then squashes the result before passing it on. What are the learned numbers and the squash step for?
• A cider-press workshop trains a grade predictor and watches a single number fall as predicted juice acidity gets closer to lab truth. What is that number?
• During one training step a foghorn-log classifier first produces a guess from the current weights, then uses the mismatch to nudge those weights. Which half is which?
• A fire-lookout hears that training “walks downhill on the error surface” by taking many small steps instead of solving for weights in one closed equation. Which idea is that?
• An herbarium model memorizes last decade’s catalog notes and then fails on this year’s new sheets, while a second model is so simple it misses obvious moisture patterns even on the training years. What are those two failures?
• A drawbridge office splits historic opening logs three ways: one pile to fit weights, one pile to choose among a few training recipes, and a locked-away pile used once at the end. What is each pile for?
• A bird-banding station spends a rainy week fitting a species tagger on archived photos, then the tagger later labels each new photo in seconds on a dock tablet. Which span is which?
• A trolley-photo archive notices that stacking layers of matrix multiplies crawls on a laptop CPU but finishes far sooner when the same work is sent to a GPU. Why does that hardware help deep training?
• A weather-balloon shed is told “the training job uses CUDA.” An intern asks whether they must write a custom GPU kernel. What should they understand?
• A lighthouse postcard archive has only a few dozen labeled captions but can start from a network already trained on a huge public text collection and adapt the last stages. What is that approach?
• A maple-sugar cooperative’s training log shows how many times the full note set was seen, how many notes were used per update, and how large each weight nudge was. Which three knobs are those?
• A swing-bridge operator can predict delay minutes from five numeric sensors, or they can draft a passenger announcement from free-text incident notes. Which model family fits which job?
• A municipal vineyard desk wants three jobs: flag blight or no blight, predict harvest tons as a number, and group unlabeled blocks of vines that behave alike. Which learning types are those?
• A flour-mill lab has mixed numeric and categorical columns and needs a strong off-the-shelf predictor. Which family fits that tabular science problem?
• A ranger first-aid post has 400 rows of lab columns and no images. One vendor sells a large language model; another offers an XGBoost model on the same sheet. Which should they choose?
• A creek-gauge office must tell the council which sensors most influenced a flood-risk score. Which method can surface that contribution?
• A marina salvage sheet has empty cells and a mix of counts and category codes. Why are boosted trees often used on that kind of table?
• A towpath museum gets four pitches: a linear model for a roughly straight fee-versus-headcount line, clustering for unlabeled visitor paths, boosted trees for a messy mix of ticket fields, and a language model for rewriting exhibit cards. How should they match those pitches?
• A county assessor hears that XGBoost can run faster on a GPU. What does that mean at associate depth?
• A lighthouse supply desk compares one shallow tree that splits on a single weather flag with a sequence of trees that each correct the previous residual. What is the second pattern?
• A harbor-pilot office wants to study which docks connect through shared tug assignments. What is the natural shape of that data?
• A volunteer trolley museum has a few hundred stops and transfer links and already works in Python on a laptop. Which toolkit fits that modest graph work?
• The same connectivity study now spans tens of millions of transfer links and must finish in a short window. Where should the graph analytics run?
• A river-freight desk asks three questions: the shortest tow route between locks, which lock is a bottleneck hub, and which clusters of landings trade mostly among themselves. Which graph-task families are those?
• A lock museum wants an assistant that can follow “this exhibit was restored by the same shop that built that engine.” Why prefer a graph of entities and relations over a single FAQ spreadsheet?
• Three GPU jobs land on one desk: wrangle a giant CSV, train a classical GPU model, and run community detection on a huge network. Which RAPIDS libraries match those jobs?
• A canal-archive team used to feed tokens one after another through a recurrent net and lost track of early tokens. Why did transformers become the building block of modern LLMs?
• A lock log says “the lower gate jammed after the upper gate opened.” The model must let “jammed” look at “lower gate,” not just nearby filler words. Which mechanism scores those relationships?
• An intern describes attention as asking a question about the current token, matching that question against other tokens, and then gathering useful content from the matches. Which three jobs are those?
• One head of a ferry-schedule model tracks who-did-what, another tracks times, and another tracks places. Why use several attention heads instead of one averaged view?
• The notices “crew relieved captain” and “captain relieved crew” would look the same if the model only saw an unordered bag of token vectors. What restores order?
• A tide-table caption is split into pieces that may be whole words or fragments before anything is embedded. What are those pieces?
• After tokenization, each piece becomes a list of numbers so that “skiff” and “dinghy” land nearer than “skiff” and “invoice.” What are those vectors?
• A lighthouse keeper pastes three seasons of logs into a chat box and the tail of the prompt is silently dropped. What limit was hit?
• Three tools sit on a docent’s bench: one reads a whole plaque and labels it, one writes the next sentence of a story, and one reads a source language then writes a target language. Which architecture families are those?
• While writing the next word of a lock announcement, a decoder must not already see the words it has not written yet. Which rule enforces that?
• A plaque-tagging model can let “closed” look at both “temporarily” on its left and “for painting” on its right. What kind of attention is that?
• After tokens have mixed information through attention, each position still goes through the same small multilayer network of its own. What is that sublayer?
• A deep stack of transformer blocks would be brittle if each block replaced its input outright. Which associate-level tricks keep that stack trainable?
• A ferry chatbot emits one piece, then feeds that piece back in to choose the following piece, and so on. What is that loop?
• During pretraining a transformer can score many token positions of a packed sequence in one pass, but at serving time a decoder still emits tokens one after another. How do those time-shapes differ?
• A bilingual exhibit card is read by one stack, then a second stack writes the other language while looking back at the first stack’s states. What is that looking-back?
• A canal-history project cannot afford librarians to tag every word in a century of minutes, yet it can hide some words and ask the model to recover them from the rest. What is that approach?
• Training blanks out a few tokens in a lock log (“The gate failed”) and scores whether the model fills the blank using both sides. Which objective is that?
• Another recipe never shows the future: the model always predicts the next word of a captain’s memoir given only the words so far. Which objective is that?
• A plaque office needs bidirectional understanding for tagging and similarity, not open-ended story writing. Which official LLM variant fits?
• A research desk hears that some modern LLMs are scaled transformer variants associated with NVIDIA’s Megatron line. How should they treat that name on this exam?
• First a model absorbs generic lock-and-harbor language from a huge unlabeled pile; later a small labeled set of “delay / no delay” notes adapts it to one desk’s ticket codes. What are those two stages?
• Early transformers still needed costly labeled corpora for each task. Why did self-supervision make BERT-, Megatron-, and similar variants so effective?
• A town council asks why a “large” language model needs huge text, many weights, and GPU time. What associate picture should they get?
• A raw next-token model completes “The lower gate” with more memoir text; after an instruction stage it instead follows “Summarize this lock log in two bullets.” What did that stage teach?
• Two variants sit on a shelf: a BERT-like encoder for tagging exhibit text and a large decoder-style Megatron-family model for drafting new copy. How should they pick?
• A hut-warden chatbot starts giving rambling answers. The team rewrites the request and adds an example instead of opening a training job. What did they change?
• First they just ask “label this incident.” Quality is uneven, so they paste three labeled incidents above the new one. What are those two patterns?
• A ferry-desk prompt is too vague, then too strict, then finally states format and audience. How should prompt work be treated?
• A museum bot should always stay in the docent voice and never invent exhibit years, while each visitor question changes. How should those pieces be split?
• A lock-fee word problem fails when the model jumps to a total, then succeeds when the prompt requires stepwise arithmetic in plain language. What prompt pattern is that?
• The same log dump must become a three-bullet brief, a list of equipment names, or a visitor Q&A. What should the three prompts do?
• A well-written instruction is buried behind an entire policy binder pasted into the prompt, and the model never sees the last chapter. What went wrong?
• A draft exhibit label is dull at a low randomness setting and wildly off-brand at a high one. What kind of controls are those?
• One failure is “tell me about the gate” with no which-gate or what-format; another is asking a small model to recite a regulation that was never in its training or prompt. What is the difference?
• A ticket office wants a bot that answers only from the pasted timetable and says when the timetable is silent. What should the prompt include?
• “Lower gate will not seat” and “the downstream leaf failed to close” should score as similar even though they share few words. Which job is that?
• A museum encodes every catalog card into a vector so a new visitor question can find the nearest cards. What is that encoding step?
• Given a lock-operation paragraph and the question “What time did the gate reopen?”, the system highlights a span already in the paragraph. What kind of QA is that?
• A new “ice-warning” ticket type appears and there is no trained head for it yet. An encoder is asked to pick among label names it was never fine-tuned on, using the meaning of those names. What is that?
• A harbor log must mark each ship name, berth, and clock time inside the sentence. Which encoder setup fits?
• Each visitor comment should become one of {praise, complaint, question}. Which encoder job is that?
• A dock office only needs a three-way tag on each radio transcript and must stay fast and on-device. Why prefer a small encoder classifier over a large generative LLM?
• After encoding, “capstan” sits near “winch” and far from “cafeteria menu.” Why does semantic search work?
• The same decoder writes a stern safety bulletin when primed with a safety template, and a cheerful kids’ tour script when primed with a tour template. What is doing the steering?
• A lock desk needs three named fields back — incident, equipment, delay minutes — not a paragraph. How should they condition the decoder?
• Staff paste a raw incident paragraph and ask for a two-sentence public notice that does not add facts. What is the source text doing?
• A visitor asks “and the Sunday one?” which only makes sense after the previous turn about winter ferries. What must condition this decoder turn?
• A decoder is asked to turn a short weather code into a spoken-style script, or a table of sailings into prose. What associate “modality” claim is fair?
• The only required output is open or closed for a gate-status feed. Should they call a long-form decoder?
• A helpful lock-bot answers every question, including how to force a gate against procedure. What gap is that?
• Reviewers rank two draft notices; those rankings later teach a preference model that nudges the generator toward the preferred style and safety. Which alignment path is that?
• A county can call an existing general LLM, adapt one with its own notices, or attempt to pretrain a new model from scratch on a tiny local pile. Which foundation-model path fits a small office?
• A timetable desk has four options: rewrite the prompt, retrieve tonight’s PDF into the prompt, fine-tune on old tickets, or pretrain a new LLM. How should they order those from lightest to heaviest?
• A clerk-bot must cite this morning’s notice, which did not exist when the LLM was trained. The system first finds relevant passages, then conditions generation on them. What is that pattern?
• Notices are encoded into vectors; a new question is encoded the same way; nearby chunks come back. What RAG loop is that?
• Sailing exceptions change twice a week. Should the desk keep fine-tuning the model on each new exception sheet, or update a retrievable store?
• The bot invents a “Tuesday ice clause” in perfect prose. What is that failure?
• A team wants an NVIDIA framework path to customize a foundation model with enterprise data, not a serving-only microservice. Which tool fits?
• The customized model is accurate but too slow on the GPU they already have. Which official tool optimizes LLM inference?
• One group wants a packaged, portable microservice for a standard model; another already serves several custom models and wants a dedicated inference server. Which NVIDIA pair matches those jobs?
• A harbor authority lists three needs in order: adapt a foundation model to local notices, then speed up token generation, then put the result behind a live endpoint. Which NVIDIA sequence is that?