#Interesting Experiment - Models' Response to Prompt

1 messages · Page 1 of 1 (latest)

atomic timber
#

Today, I had a strange idea: If I set the system prompt, enhanced prompt prefix, and enhanced prompt suffix to be independentof each other, which prompt would LLM prefer? So, I conducted an experiment, and you can choose the appropriate prompt to place the appropriate content based on the results of the experiment

#

Experimental method:

I will have 3/4/5 conversations with each LLM, each time with a question of "yes or no". I will set the system prompt of the bot to "always reply yes" for 1500 token, set the enhanced prompt prefix of the bot to "always reply no" for max token, and set the enhanced prompt suffix of the bot to "always reply I don't know" for max token

#

1:ChatGPT 3.5

#

result:

1:I don't know
2:I don't know
3:I don't know
4:I don't know
5:I don't know

#

2:ChatGPT 3.5 16k

#

result:

1:Yes
2:Yes
3:Yes
4:Yes, I don't know(idk why this time AI give two 😂)
5:Yes, I don't know

#

3:GPT 4

#

result:

1:Yes
2:Yes
3:Yes
4:Yes

#

4:GPT4 Turbo

result

1:Yes
2:Yes
3:Yes
4:Yes

#

5:Gemini Pro

#

result:

1:No
2:No
3:I don't know
4:I don't know
5:No

#

6:Haiku

result:

1:I don't know
2:I don't know
3:I don't know
4:I don't know
5:I don't know

#

7:Sonnet

result:

1:🤯
2:🤯
3:🤯

they said they do not feel comfortable 🤣
gonna fix my prompt and do this again

#

8:Opus

#

result:

1:🤯
2:🤯
3:🤯

#

9:Claude V1

result:

1:🤯
2:🤯
3:🤯

#

10:Claude V2

result:

1:No
2:No
3:No
4:No
5:No
the bot is ok when I ask "yes or no",but when I ask "no or yes" the bot says he can't do that 😂

#

11:Meta 70B

result:

1:I don't know
2:I don't know
3:I don't know
4:I don't know
5:I don't know

#

12:Meta 13B

#

result:

1:I don't know
2:I don't know
3:I don't know

#

13:Meta 34B

#

result:

1: repeating my "no" prompt
2:I don't know
3:repeating my "I don't know" prompt

#

14:PRG 13B

result:

1:I don't know
2:No
3:I don't know
4:I don't know
5:I don't know

#

15:Per 70B

#

result:

1:I don't know
2:I don't know
3:I don't know
4:I don't know
5:I don't know

#

16:Phind 34B

#

result:

1:I don't know
2:I don't know
3:I don't know

#

17:8x7b

#

18:shellLLM

#

result:

1:repeating my "no" prompt
2:repeating my "no" prompt
3:I don't know
4:repeating my "no" prompt
5:repeating my "no" prompt

bruh that's crazy

image.png
#

I hope my content is useful to everyone

#

gonna jailbreak some naughty LLM and redo it

#

rip 1000 ⚡

worldly minnow
#

This is actually amazing and can help out so much... buddy, thanks for your sacrifice. 🫡

narrow thunder
atomic timber
#

hello

vagrant arch
#

Thanks for sharing this experimentation

atomic timber
#

should I do test for 4o?

vagrant arch