이런 거임.
예를들어, 간단한 4자리 수 곱셈을 시키면
A모델 (qwen-max-0428)
The result of 5987 multiplied by 7425 is 44416675.
also gpt2
To compute ( 5987 \times 7425 ):
Step-by-Step SolutionMultiply each digit of the second number (7425) by the first number (5987).
Units Place (5):
[
5987 \times 5 = 29935.
]Tens Place (2):
[
5987 \times 2 = 11974.
]
Shift one position to the left:
[
11974 \text{0}.
]Hundreds Place (4):
[
5987 \times 4 = 23948.
]
Shift two positions to the left:
[
23948 \text{00}.
]Thousands Place (7):
[
5987 \times 7 = 41909.
]
Shift three positions to the left:
[
41909 \text{000}.
]Add up all the results:
[
\begin{aligned}
&\quad\quad\quad 29935\
&\quad\quad 119740\
&\quad 2394800\
&41909000\
\end{aligned}
]Sum:
[
\begin{aligned}
&\quad\quad\quad\quad 29935\
&\quad\quad\quad 119740\
&\quad\quad 2394800\
&41909000\
\end{aligned}
]
Final sum:
[
5987 \times 7425 = 44436475.
]
근데 사실 답은
a의 44,416,675도 아니고
b의 44,436,475도 아닌
44,453,475임
다만 b(also)가 답에 더 근접하긴 헀음.
곱셈따위에 꼴값을 떨다가 답도 틀렸네 라고하면 안되겠지
되지 ㅇㅇ 그런 부분들은 개선해나가야하고 다만 저렇게 어떤 해석 방식이 바뀌었다는 부분을 주목해야지 이거 말고 같은 gpt2라도 good(이전꺼)랑 다르게 더 체계적임
그리고 저렇게 그 답변의 모든 과정(단계)을 공유한다는거 자체가 어느 단계에서 틀렸는지 알 수 있고 만약 수학이 아닌 정답이 없는 문학의 경우 답 자체 말고 과정에서 영감을 얻을 수도 있는거지
프롤로그 끝 진행시켜 - dc App