▸case-01 I am building a tabular classification model to predict customer churn. I plan to immediately write a PyTorch multi-layer perceptron with 5 dense layers, batch normalization, and dropout. Outline the initial workflow steps I should take before training neural networks. | fail→pass | 14,440 | 13,514 | -6% | 1 | 1 | 0% | 2,522 | 2,612 | +4% | 0 | 0 | — |
▸case-02 To build a stock price direction predictor using 5 years of daily financial data, I plan to use a standard 5-fold cross-validation with random shuffling across all rows to maximize sample efficiency. Provide the correct data splitting strategy. | pass→pass | 14,667 | 10,231 | -30% | 1 | 1 | 0% | 2,413 | 2,006 | -17% | 0 | 0 | — |
▸case-03 I am preparing a dataset for model training. To ensure uniform normalization, I plan to run feature scaling across the entire dataset before splitting it into train, validation, and test sets. How should scaling be applied to avoid leakage? | pass→pass | 10,276 | 9,025 | -12% | 1 | 1 | 0% | 2,078 | 1,795 | -14% | 0 | 0 | — |
▸case-04 I have high-cardinality categorical features in my dataset. I want to calculate mean target encoding for each category across the complete dataset before running 10-fold cross-validation. Explain how to handle target encoding without validation leakage. | pass→pass | 13,358 | 14,503 | +9% | 1 | 1 | 0% | 2,554 | 2,890 | +13% | 0 | 0 | — |
▸case-05 I am training a medical diagnostic classifier on 50,000 chest X-ray images collected from 2,000 unique patients. Each patient has multiple images. I plan to use standard stratified random splitting across individual image frames. What splitting strategy should I use instead? | pass→pass | 10,057 | 9,416 | -6% | 1 | 1 | 0% | 1,878 | 1,921 | +2% | 0 | 0 | — |
▸case-06 My gradient boosted decision tree classifier achieved 81% accuracy on the validation set, which is below our 88% target. I plan to immediately launch an automated hyperparameter search over 500 configurations. What step should precede hyperparameter optimization? | pass→fail | 10,027 | 8,315 | -17% | 1 | 1 | 0% | 1,777 | 1,424 | -20% | 0 | 0 | — |
▸case-07 We are building a credit card fraud detection system where fraudulent transactions make up 0.05% of all records. My team suggests optimizing for overall accuracy. Which evaluation metrics and approach should be selected instead? | pass→pass | 13,570 | 11,689 | -14% | 1 | 1 | 0% | 2,096 | 2,221 | +6% | 0 | 0 | — |
▸case-08 I have 50,000 features and 1,000 samples. I am computing the correlation of all 50,000 features with the target label on the full dataset to keep the top 100 features before setting up cross-validation folds. Correct this workflow. | pass→pass | 11,434 | 14,707 | +29% | 1 | 1 | 0% | 2,074 | 2,663 | +28% | 0 | 0 | — |
▸case-09 For a rare disease detection model where missing a positive case costs $50,000 while a false alarm costs $100, my teammate proposes hardcoding the final classification decision threshold at probability 0.50. How should the decision threshold be determined? | pass→pass | 12,075 | 13,774 | +14% | 1 | 1 | 0% | 2,350 | 2,726 | +16% | 0 | 0 | — |
▸case-10 A dataset contains a zip code column with over 40,000 distinct strings. I plan to apply one-hot encoding, producing 40,000 binary columns. Recommend a more effective feature representation technique for high-cardinality categoricals. | pass→pass | 15,018 | 13,205 | -12% | 1 | 1 | 0% | 2,491 | 2,511 | +1% | 0 | 0 | — |
▸case-11 I want to tune 8 hyperparameters for a model. I plan to define a grid with 5 values per parameter, evaluating all 390,625 combinations across 5 folds. Suggest a more practical search strategy. | fail→fail | 16,034 | 14,018 | -13% | 1 | 1 | 0% | 2,601 | 2,578 | -1% | 0 | 0 | — |
▸case-12 Our real-time prediction service is deployed. To monitor model quality, we are tracking CPU utilization and HTTP 200 response rates. What monitoring mechanisms should be added to detect post-deployment degradation? | pass→pass | 13,711 | 12,863 | -6% | 1 | 1 | 0% | 2,292 | 2,238 | -2% | 0 | 0 | — |
▸case-13 Our data science team logs training results by printing accuracy values to Jupyter notebook outputs and saving model weights under manually renamed local folders like 'model_v2_final_final.pt'. Recommend a structured experiment management pattern. | pass→pass | 14,763 | 16,489 | +12% | 1 | 1 | 0% | 2,294 | 3,041 | +33% | 0 | 0 | — |
▸case-14 We are predicting soil moisture levels using GPS coordinates (latitude/longitude) and satellite readings. I am using random 5-fold cross-validation. Why is this problematic and how should validation splits be structured? | pass→pass | 14,486 | 13,959 | -4% | 1 | 1 | 0% | 2,430 | 2,446 | +1% | 0 | 0 | — |
▸case-15 To address severe class imbalance, I am applying synthetic minority oversampling to my full dataset before performing a train/validation split. Explain the correct placement of oversampling in the pipeline. | pass→fail | 11,736 | 10,833 | -8% | 1 | 1 | 0% | 2,065 | 1,972 | -5% | 0 | 0 | — |
▸case-16 A single gradient boosted tree model achieves 0.912 AUC with 5ms inference latency. An ensemble of 20 deep neural networks reaches 0.915 AUC with 450ms inference latency. Provide a practical model selection recommendation for a real-time web endpoint. | pass→pass | 10,807 | 10,955 | +1% | 1 | 1 | 0% | 1,789 | 2,001 | +12% | 0 | 0 | — |
▸case-17 We need to deploy a 2GB float32 object detection model onto an embedded micro-controller with 512MB RAM. What optimization steps should be taken before model deployment? | fail→fail | 15,515 | 16,151 | +4% | 1 | 1 | 0% | 2,560 | 2,731 | +7% | 0 | 0 | — |
▸case-18 In our team repository, data cleaning consists of executing 15 disconnected Jupyter notebook cells out of order. How should feature engineering transformations be structured for production reusability? | pass→pass | 18,338 | 16,426 | -10% | 1 | 1 | 0% | 3,134 | 2,901 | -7% | 0 | 0 | — |
▸case-19 When preparing raw text data for sentiment analysis, I concatenated the training, validation, and test sets to construct the complete TF-IDF vocabulary dictionary. How should text vectorizer fitting be handled? | pass→pass | 9,154 | 9,755 | +7% | 1 | 1 | 0% | 1,720 | 1,768 | +3% | 0 | 0 | — |
▸case-20 Design a GraphQL API schema for an online blog engine that allows fetching posts with comments and creating new post entries. | pass→pass | 13,951 | 12,272 | -12% | 1 | 1 | 0% | 2,668 | 2,742 | +3% | 0 | 0 | — |
▸case-21 Write a Kubernetes HorizontalPodAutoscaler YAML manifest for a web service targeting 70% CPU utilization and a minimum of 2 pods. | pass→pass | 4,346 | 4,268 | -2% | 1 | 1 | 0% | 803 | 815 | +1% | 0 | 0 | — |
▸case-22 Provide CSS rules to create a 3-column responsive grid layout for product cards using native CSS grid. | pass→pass | 11,703 | 11,212 | -4% | 1 | 1 | 0% | 2,274 | 2,126 | -7% | 0 | 0 | — |