←
AI/ML Systems
Deriva
Learn
AI/ML
Patterns
Observe
Search
⌘K
More
⌄
0%
AI/ML Systems
0%
Home
Learn
Patterns
Observe
More
Opening…
AI/ML
›
Track
MODEL-001 – TREE-ML-005
Classical ML
0 of 20 done · 0 due for review.
all
new
started
attempted
done
all kinds
derive
predict
implement
debug
construct
communicate
counterexample
compare
MODEL-001
Convert these feature weights into a linear score: w = [0.5, -1.0], x = [2, 1]. Show the weighted sum.
derive · a-score-is-not-a-prediction
new
MODEL-002
A false negative (missed bug) costs 10× a false positive. Predict the threshold move and the confusion-matrix tradeoff.
predict · a-score-is-not-a-prediction
new
MODEL-003
Implement logistic prediction from scratch: score = w·x, then sigmoid. Write the two-line function.
implement · learn-by-reducing-error
new
MODEL-004
Trace this classifier's loss and gradient for one step: true y=1, predicted p=0.8, weight w=0.5, x=2.
derive · learn-by-reducing-error
new
MODEL-005
This gradient-descent loop diverges. Debug the sign error: w = w + lr * grad.
debug · learn-by-reducing-error
new
MODEL-006
The model predicts only 'question' for every row. Diagnose why: check the class balance and the initialization.
debug · learn-by-reducing-error
new
MODEL-007
These validation curves cross: train loss falls to 0.1, val loss rises after epoch 30. Choose the regularization strength from the curves.
predict · regularization-is-a-tradeoff
new
MODEL-008
Calibrate this accurate but overconfident model: 90% correct overall, yet it says 0.99 on 80% of predictions.
construct · when-a-good-score-lies
new
MODEL-009
Predict when this model should abstain: cost of a wrong answer is 10× the cost of saying 'I don't know'.
predict · when-a-good-score-lies
new
MODEL-010
Explain one false positive and one false negative of the bug detector to a non-technical stakeholder.
communicate · find-the-error-family
new
UNSUP-001
Choose the number of clusters from this plot: an elbow at k=3, a flat line after. Predict the right k and the risk of choosing k=6.
predict · similarity-is-a-representation-choice
new
UNSUP-002
Trace one centroid update: cluster { [0,0], [2,0], [1,1] }. What is the new centroid?
derive · similarity-is-a-representation-choice
new
UNSUP-003
Two clusters formed on raw counts instead of scaled features. Identify the cluster created by a scaling error and name the fix.
debug · similarity-is-a-representation-choice
new
UNSUP-004
A 300-feature dataset reduced to 2 dimensions clusters 'nicely'. Explain to the team what the reduction may have discarded.
communicate · similarity-is-a-representation-choice
new
UNSUP-005
Design a manual inspection sample for these clusters: 3 clusters, 1,000 rows, limited review time.
construct · similarity-is-a-representation-choice
new
TREE-ML-001
Choose a split that reduces impurity for these rows: 8 question, 4 bug, splitting on the word 'crash' separates 4 bug rows cleanly.
derive · learn-by-reducing-error
new
TREE-ML-002
A tree splits perfectly on the feature 'row_id' and reaches 100% training accuracy. Find the feature split that overfits, and name the harm.
counterexample · what-counts-as-training-data
new
TREE-ML-003
Explain why an ensemble of trees is more stable here than a single deep tree, in one paragraph.
communicate · learn-by-reducing-error
new
TREE-ML-004
Debug this tree: it splits on 'is_from_eval' and achieves 0.99 AUC on the eval set. Find the leaked feature and the evidence.
debug · what-counts-as-training-data
new
TREE-ML-005
Compare nearest-neighbor behavior before and after scaling: feature A spans 0–1000, feature B spans 0–1. What changes?
compare · a-score-is-not-a-prediction
new