#Even a model at zero temperature can
carmine needle
The whole idea of graph of thoughts is to generate multiple thoughts for the same prompt no? I'm just confused on how it would always generate independent thoughts
finite coral
I understand this question. The context will be different. Even the time stamp tokens contribute.
My custom GPTs use a similar technique to graph of thoughts, called chain of reason.
carmine needle
How would the context be different though if it's the same prompt??
finite coral
they're identified as different agents with different roles, and each comments on the context in turn, building or refuting an idea.
finite coral
How would the context be different though if it's the same prompt??
even the timestamps influence something as nuanced as the instruction "improvise".
Zero temp is going to improvise poorly, but it will improvise.
At least it should. You can test it out fairly cheaply with 3.5 model, I'd guess.
carmine needle
what do you mean by the instruction improvise??
Ok so basically though, it's better to have temp > 0 right
closer to 1
finite coral
what do you mean by the instruction improvise??
if your prompt tells the AI to use tree of thought, it's improvising from the context, even if you don't use that word.
improvisation is the practice of making up the next step from the current context.
carmine needle
So, right here, it generates 10 responses. When I tried the same prompt with Claude2. it was giving me the same responses over and over. no variety
finite coral
it's literally a type of inference in human minds. but AI will do this. include timestamps in the prompt at zero temp for variatrion, is what i'm explaining
carmine needle
it's from the paper
carmine needle
Hmm so I don't explicitly tell my LLM to use grpah of thoughts in the prompts. I'm just trying it with the code on the paper
finite coral
okay, that wasn't clear.
carmine needle
and I see that they all set the temp parameter to .6 or 1
also thanks alot by the way for explaining. it's really helping
finite coral
0.6 is great for this kind of thing
it gives room for originality without needing to instruct for it
finite coral
also thanks alot by the way for explaining. it's really helping
no worries.
carmine needle
okay but let's say I don't give it 0.6. I give 0. Then I should say use "graph of thoughts" in the prompt??
finite coral
in terms of getting different output, it's never the same time. pass that to the model with the prompt and even zero temp gives unique output when instructed to include improvised output.
carmine needle
in terms of getting different output, it's *never* the same time. pass that to t...
Yes okay. so what would this look like? that's what I'm curious about
finite coral
okay but let's say I don't give it 0.6. I give 0. Then I should say use "graph o...
not quite that simple. you'll have to provide explicit instructions for how to do it and an output template for reliable compliance.
You'd just pass the timestamp with your prompt in the API payload.
I don't really work with the API, but if I were doing this, I'd modify the client.
I have a visual disability though that makes Python physically uncomfortable (extremely) to work with, so I stick to the chat interface and prompt at 0.7 temp 😄
carmine needle
okay i see that makes sense
finite coral
But if you passed the time/date stamp with a prompt, you can expect every single one will be distinct. That could provoke different output.
I digress. In your case you probably have a conflict in your system prompt/instructions/context.
i can't find it for you, but i can give you tips on what to look for:
- logical inconsistencies like quantities and conditions.
- negative instructions. these will get interpreted as positive on a long enough context.
- poorly defined or aligned instructions. this often involves tasks that aren't computable. see number 1.
- you need to use a good output template to get reliable compliance from any prompt strategy, be it ReAct, Tree of Thoughts, what have you.
I hope this is helpful and gets your project on track.
carmine needle
MULTI_DF_PREFIX = """
You are working with 3 pandas dataframes in Python named df_app_anon, df_db_anon, df_infra_anon. You
should use the tools below to answer the question posed of you:"""
SUFFIX_WITH_DF = """
If you have already performed an action, do not repeat it, try a different action.
If you encounter UndefinedVariableError, then only query on the provided column names below.
Do not modify the dataframe in any manner, only query operations are allowed.
This is the result of column names of all dataframes and their associated data types.
{df_col_and_dtypes}
Begin!
Question: {input}
{agent_scratchpad}"""
This is my prompt
SO I do give negative instructions