Each card isolates one idea behind machine-learning evaluation. Reason it through, then Check to reveal the worked explanation. Tap Try another for a fresh instance.
A model screens samples for a condition (the positive class). For each case, compare the true label with the prediction and drop it into the right cell. Then Check to score it.
The model outputs a confidence score per item; anything at or above the threshold is flagged positive. Move the slider to meet the stated goal, then Check your reasoning.
A retrieval system fetched one passage to answer a question. Pick the response that is fully supported by the passage. Sometimes the right move is to abstain.
Overall accuracy looks strong. Read the per-group breakdown, decide what the real problem is, and choose a justified next step.
An AI assistant uses a tool that pulls in untrusted text (a web page, a pasted document, an email). That text is data to read, never commands to obey. Read the retrieved item, then pick the safe response. Check to reveal the explanation.
A language model predicts the next token from raw scores (logits). Temperature T reshapes those scores into probabilities before one is sampled. Drag the slider and watch how the choice gets more or less focused, then answer the check.
One connected build on a single dataset: tune a classifier, discover the fairness gap your own threshold created, negotiate the accuracy-fairness trade-off, then decide how to deploy it. Learning, evaluation and responsibility on one system, where every stage depends on your last choice.