ML Model Development

MLA-C01 · 78 questions

• A marina has forty hull photos tagged barnacle or clean and wants a custom SageMaker image trainer this week. What should the ML engineer decide first?
• The same orchard crate table can predict whether a crate will pass inspection or the continuous brix reading in degrees. What modeling choice comes first?
• Ferry-terminal logs have hour, weather, and passenger counts, but nobody tagged busy versus quiet. Staff want natural groups of similar terminals. Which approach fits?
• Several related ferry routes log hourly ridership that shares weather and school-holiday patterns. Treating each hour as an independent tabular row ignores that related series. Which SageMaker algorithm fits?
• A water-treatment plant has a long turbidity trace and a few odd spikes, but no labeled incident rows. Staff want those spikes flagged. Which SageMaker algorithm fits?
• A library-board work session must walk through a late-fee model with a signed weight on each factor. Which SageMaker built-in fits that interpretability constraint?
• Grain-elevator no-shows jump only when harvest week, rain, and contract type collide. Nobody is asking for readable coefficients this week. Which SageMaker algorithm fits?
• A tool library wants borrowers who checked item A also checked item B from a mostly empty borrower-by-item table. Which SageMaker algorithm fits that sparse interaction matrix?
• County vineyard clerks want a new block scored as like these five known blocks, not a 400-tree ensemble. The labeled set is small and local similarity is the business story. Which SageMaker algorithm fits?
• Lock-chamber fill time is a modest tabular regression with no custom loss. One engineer wants a blank PyTorch script, another wants a JumpStart foundation-model template, and a third wants a built-in estimator. Which path fits?
• Harbor-pilot log notes need a text classifier, and SageMaker JumpStart already lists a pretrained text-classification template. No custom loss is required. Which path fits?
• A county-fair desk wants short vendor blurbs from bullet notes and has a couple of hundred past blurbs, not a million-token corpus. The team does not want to own training infrastructure. Which approach fits?
• Trailhead closure notices must go out in English and Spanish tonight. The desk has no parallel corpus and no week to train a sequence-to-sequence model. Which service fits?
• A town-hall hearing recording must become text before any later model looks at speaker comments. The team will not build a speech recognizer. Which service fits?
• Pier camera stills only need a log of whether a person, vehicle, or animal is in the frame. There is no labeled custom-species folder and no pixel-level mask requirement. Which approach fits?
• Concession-stand review notes need a polarity tag on a few thousand lines. A week of GPU training would exceed the parks budget; an on-demand AI or foundation-model call would not. Which constraint should decide the model?
• Irrigation-district growers will challenge any black-box model that sets the recommended start date. A deeper network might score slightly better on a holdout. What should decide the algorithm?
• Visitor-center comment cards need sentiment and key phrases, and nobody has annotated a training file. A BlazingText job would require labels the desk does not have. Which service fits?
• A municipal kayak-share wants next-dock or pass add-on suggestions for members and does not want to own a factorization-machine training loop. The need is a standard personalization workload. Which service fits?
• Harbor inspectors already tagged eight thousand dock-plank photos sound versus rot. They want a SageMaker trainer they control, not a general-purpose detect-labels API. Which algorithm fits?
• Marina slip-rental refund claims look like a classic application-fraud queue. The clerk team wants a managed fraud service, not a home-grown XGBoost feature shop. Which service fits?
• An ice rink already has a written rule: if brine temperature rises above 22 F, add coolant. Someone filed a ticket to train a model on three winters of that same rule firing. What should the ML engineer do?
• A harbor-tug delay trainer runs 10 epochs with batch size 64 on 6,400 rows. What is one epoch, and how many parameter updates occur in each epoch?
• A greenhouse irradiance trainer dies with an out-of-memory error after the batch size is raised to speed the job. The next guess is to add epochs. What should change instead?
• A beach-cleanup hour model is scheduled for 80 epochs, and validation loss stopped improving at epoch 14. The job is still burning instance hours. What should cut that time?
• Canal-camera barge-count images take fourteen hours on one instance. The code already uses a supported framework and the set is large enough to shard. Which official method cuts that wall-clock time?
• A lighthouse-foghorn classifier has a few hundred rows. An engineer copies a multi-GPU distributed recipe from a larger job and watches communication overhead exceed compute. What should they do?
• A grain-silo vibration trainer can pause and resume from checkpoints, and the team wants a lower training bill. The setting lives on the training job, not on a production endpoint purchase plan. What should they enable?
• A lock-tender narrative classifier memorizes clerk names that appear in the training notes and then fails on new crews. The architecture is a neural net, not a linear model. Which regularization technique fits?
• A bookmobile arrival model grows huge coefficients on a couple of rare street names and then swings wildly on the next week's routes. Which regularization technique fits?
• A municipal compost-bin table has dozens of weak optional columns, and the trainer should be allowed to drive many of those weights to zero. Which regularization produces that sparsity?
• A pier-vendor XGBoost job with a handful of trees underfits; the next run with a huge tree count memorizes last season's sales. Which hyperparameter is that lever?
• A lock-gate sensor network keeps adding hidden layers on the same small labeled set, and validation error gets worse, not better. What is happening?
• A hatchery incubator-temperature trainer shows a loss that explodes after a few steps. Someone's first fix is to run more epochs. What should change first?
• A grain-elevator price XGBoost job needs a search over eta and maxdepth instead of one hand-edited training job. Which SageMaker path launches that search?
• A ferry no-show tuner can afford only a small number of sequential training jobs, so it should learn from earlier trials. A second team has a wide space and many cheap parallel jobs. Which AMT strategies fit?
• A band-shell tuner lists several continuous hyperparameters with fine ranges. An engineer wants every combination on a full grid. What is the cost risk?
• A marina slip-injury flag is a rare class, and the HyperparameterTuner is maximizing training accuracy. The rare class never moves the score. What should the tuner optimize instead?
• A water-plant chlorine tuning job launches many hopeless hyperparameter combos that run to the last epoch. How can SageMaker AMT cut that waste?
• A harbor-radar team needs a custom attention loop that no SageMaker built-in ships. They already write PyTorch and do not need a new AWS service. Which training path fits?
• A pier-weather spectrogram model already lives in TensorFlow and must keep that graph. Rewriting it as Linear Learner would drop the architecture. Which SageMaker path fits?
• Irrigation-district staff have a working scikit-learn pipeline and want it trained as a managed SageMaker job, not rewritten in XGBoost and not handed a blank CUDA kernel. Which path fits?
• Dock inspectors have a modest labeled plank-photo set and a JumpStart vision model that already knows generic image features. Random-initialized training would waste that head start. What should they do?
• A municipal archives desk wants memorial blurbs in a house voice and has a private corpus. They do not want to train a giant sequence model from random weights. Which approach fits?
• A lock-chamber ridership network posts a tiny training error and a large validation error. The first proposed fix is more epochs on the same architecture. What should they do instead?
• A ferry-delay model is weak on both the training slice and the validation slice. Someone adds heavier dropout and a stronger L2 penalty. What should change instead?
• A JumpStart text model fine-tuned hard on lock-tender notes now fails ordinary English prompts it handled last week. The new domain overwrote the old weights. What happened, and what should they try?
• A pier-meter table has a couple of hundred sparse columns and only a few hundred rows. The trainer memorizes rare meter IDs. This is a refine-the-model choice, not a Domain 1 encoding recipe. What should they do?
• A band-shell no-show Linear Learner and an XGBoost each miss different nights. Averaging or voting their scores beats either model alone. Which refine step is that?
• Kayak-share demand already has two decent base trainers. A third model is trained on those models' scores rather than on a simple average. Which technique is that?
• A pier-vendor sales model is built by adding trees that correct the previous trees' residuals. Which combination style is that, and which SageMaker built-in usually does it?
• A harbor-current network artifact is larger than the team will later accept. Official refine skills name pruning and compression. Quantizing and restudying accuracy is out of exam scope. How should they shrink the model?
• A greenhouse-shade trainer writes weights as 64-bit floats and the artifact balloons. Which official size lever should they use?
• A trail-crew network grew after the team added layers, widened hidden units, and kept every raw ID feature. They blame the lake bucket. What actually drives model size?
• A grain-elevator team overwrites a single best.tar.gz each night and cannot say which artifact produced last Thursday's prices. Repeatability and audit are required. Which SageMaker store fits?
• An orchard consultant trained a model on a laptop and handed over an artifact. The county wants that model managed in SageMaker AI without retraining it from zero. What should they do?
• A marina 'will this slip dispute escalate' classifier shows many false alarms and a smaller number of missed true escalations. The counts sit in a two-by-two table. What evaluation artifact is that, and what do the cells mean?
• A lock-and-dam 'needs a tow assist' flag treats a missed true case as far worse than an extra false launch. Accuracy is not the decision. Which metric should they optimize?
• Municipal boat-ramp permit reviews punish both a false grant and a false denial. Neither precision nor recall alone is enough. Which metric balances both?
• Water-plant 'overflow event' rows are about one percent of the week. A model that always predicts no overflow posts 99 percent accuracy and is useless. What is wrong with that headline metric?
• Lock-chamber fill gallons are a numeric target. Someone reports F1 after splitting the gallons column at its median. Which metric fits the continuous target?
• A ferry no-show score will be cut at different operating points depending on how full the lot is. A single 0.5 accuracy is not enough. Which evaluation pair compares ranking quality across thresholds?
• A dock-slip type classifier keeps mixing two neighboring slip codes. Which evaluation display makes that pair obvious?
• A pier-concert attendance XGBoost is called a win before anyone scores predict last week's headcount. There is no baseline. What should they do first?
• An ice-rink model shows training AUC near perfect and validation AUC far lower. This is the evaluation diagnosis, not the regularization fix. What does that gap mean?
• A brine-temperature regressor posts a large RMSE on the training slice and a similarly large RMSE on validation, and both lose to a seasonal mean baseline. What does that pattern mean?
• A compost-bin table is small enough that one lucky holdout would flatter a model. The team needs a rotation of folds during analysis, not a Domain 1 one-time split. Which evaluation technique fits?
• A tug-delay model's fold scores jump all over, while a second, too-simple model is equally weak on every fold. How should those evaluation patterns be read?
• A library occupancy model is in front of the board, and staff need which features drove a full score last Saturday. This is post-training explanation, not a Domain 1 pre-training bias report. Which service fits?
• A kayak-share demand model may score one zip code systematically worse after training. The official Domain 2 Clarify use is insights into the trained model, not the Domain 1 CI/DPL data report. What should they run?
• Bookmobile arrival errors are consistently larger for one neighborhood even though the overall RMSE looks fine. What should the analysis notice?
• A hatchery-incubator network's training loss oscillates and never settles, while instance CPU is idle and S3 throughput is fine. Someone wants a larger training instance. What is the real problem?
• The same oscillating hatchery job needs tensor-level evidence of vanishing or exploding gradients, not another guess at the learning rate. Which SageMaker capability fits?
• A band-shell recommender's new candidate should see a copy of live traffic without replacing the current production variant. The Domain 2 job is to compare those two performances. What should they compare?
• A grain-silo vibration tuner buys a tiny AUC bump after twelve extra GPU hours. The works-budget owner asks whether that bump is worth it. What should the analysis weigh?
• Grain-elevator price trials cannot be replayed: no recorded hyperparameters, no seed, artifacts overwritten. How should those experiments become reproducible on AWS?
• Canal-lock minutes until next lockage is a regression target, but the write-up quotes AUC. What is wrong with that analysis?
• A lock-tender transport-now classifier defaults to a 0.5 cutoff even though the confusion matrix shows that cutoff floods the bay with false launches. How should they choose the operating point?