dumb question but "why can't we just" 🙈 #70
Replies: 1 comment 4 replies
-
Defeats the purpose of the project, which is to demonstrate that the o1 reasoning advance is replicable. Swiping their CoTs will not accomplish that goal. |
Beta Was this translation helpful? Give feedback.
-
|
No, I'm no suggesting you swipe O1's CoT - I'm suggesting you could take any existing open source implementation of a CoT or ToT agent (backed by any chat model with an appropriate license) and modify it slightly to capture and annotate it's internal input/output and train on that. |
Beta Was this translation helpful? Give feedback.
-
ToT requires code for proper execution so I assume what you mean is we're combining every intermediate generation and the final answer into what will be treated as the full answer during finetuning. this is an idea I had ages ago, even before Q* was first referenced, in that time I figured "Graph of Thought" is a much better option and the inclusion of "Self Discover" could further diversify thought and improve the final output. I truly hope david comes back to read this |
Beta Was this translation helpful? Give feedback.
-
That is what I'm trying to say, yes, thank you. 🙂 |
Beta Was this translation helpful? Give feedback.
-
|
Someone mentioned this elsewhere: https://github.com/DataBassGit/o7 "Agent framework for generating a synthetic dataset" 👀 |
Beta Was this translation helpful? Give feedback.
-
This suggestion is way too obvious to work, but humor me:
Why can't we just take the input/output from an existing CoT or ToT agent program and fine-tune on that?
Literally just run queries through the most successful CoT/ToT agent and train on their output?
We already have programs that produce the kind of reasoning workflow we'd like the LLM to learn, right?
Obviously someone has thought of this and it won't work, I'm just curious to learn why. 😌
Beta Was this translation helpful? Give feedback.
All reactions