問題文
A team is starting work on a product that must answer questions about photographs of retail shelves. Their first plan is to train a large image-and-text model from scratch on eight weeks of collected photos. A senior engineer asks them to establish a baseline first. What is the strongest reason for running a baseline before the large training job?
選択肢
- A baseline guarantees that the larger model will converge, and the smaller run confirms the loss function is correct.
- A baseline model trained on a small subset will always reach the same accuracy as the full model, because the evaluation set is what determines the score and the amount of training data only changes how long the run takes to finish.
- A baseline gives a reference number on the same evaluation set, so any later gain can be attributed to a specific change rather than to the extra effort.
- A baseline removes the need for a held-out evaluation set, and the comparison between two runs is itself a form of validation.