問題文
Two features are being tuned. Feature A transcribes and reformats a spoken instruction into a fixed template. Feature B writes three alternative marketing headlines for a product. Which decoding setup suits each?
選択肢
- Deterministic decoding for both, as a reproducible output is required in every production setting.
- Sampling for A, and drawing from the distribution stops the model getting stuck on a phrase from the template.
- Deterministic decoding for A, since one correct rendering is wanted, and sampling for B, since variety among the three outputs is the point.
- Beam search for both, since it always produces the highest-quality output, for the template and the headlines alike.