フリー問題

問題文

A classification prompt scores 84 percent with 8 examples in the request and 85 percent with 32 examples. Each request now carries 4,000 extra tokens. What is the appropriate reading of this result?

選択肢

1. The returns from adding examples have flattened, so the extra cost per request is no longer buying meaningful accuracy.
2. Accuracy will keep rising roughly linearly, so the team should try 128 examples next.
3. The result shows that the model has memorized the 8 examples, so the team should shuffle them on every request.
4. The 1-point gain proves the prompt is now optimal and needs no further work.