Test de instruire AI

Răspunsurile și explicațiile apar pe măsură ce susțineți testul.

1. Care intrare din jurnalul de antrenament descrie o epocă completă?
   • O trecere completă prin setul de date de antrenament
   • O actualizare a optimizerului dintr-un singur exemplu
   • Un test efectuat după implementare
   • Un înregistrare a unei predicții a modelului
2. Which sequence describes a basic supervised training step?
   • The test set updates weights, then training labels are generated
   • The model writes a target answer, then deletes the input
   • The model predicts, loss compares with a target, and an optimizer updates parameters
   • A prompt changes weights and the loss is ignored
3. Validation has selected a training configuration. Why evaluate that model on a separate test set next?
   • To use its labels for every weight update
   • To evaluate the selected model on data not used to fit or choose it
   • To make the training loss smaller
   • To store preprocessing rules
4. How can a data split leak information?
   • All records from one source are kept together
   • Later periods are used only for later forecasts
   • Preprocessing rules are stored with the model
   • Related records from the same person or event are split across train and test
5. Does typing a prompt into an already trained model necessarily train it?
   • No; a prompt changes the current context, while weight updates are a separate process
   • Yes; every prompt immediately updates model parameters
   • Yes, but only if the answer is long
   • No; prompts are ignored by trained models
6. What should be saved with a training checkpoint to support reproducibility?
   • Only the final model filename
   • Only the training loss
   • Data version, preprocessing rules, model configuration, and evaluation results
   • The test labels mixed into training
7. Why does the training objective matter?
   • It guarantees the model learns the outcome that users care about
   • It determines what the model is rewarded for learning
   • It is the same as the model’s deployment policy
   • It removes the need for test data
8. What does a lower training loss establish?
   • That the model will generalize to all new examples
   • That the labels are factually correct
   • That deployment monitoring is unnecessary
   • Better agreement with the training objective, not reliability on every future input