How to contribute data.
Closed this issue · 1 comments
There are some questions/tasks I'd like an open-source reasoning model to be able to answer, is there some way I can contribute those to this dataset?
Also, are you going to do synthetic generation, and if so, what's the process to participate in that?
Make suggestions/questions here regarding data:
https://github.com/daveshap/Raspberry/discussions/categories/data
We aren't quite at the synthesis phase yet. We're still gathering ideas. So far:
- CoT: yes, we all tend agree this is the central concept
- Reflection: Probably
- MCTS: this is being talked about a bit
Then, as far as actually synthesizing data for reasoning, we're talking about using provable tests for synthesis (e.g. something that allows us to provably verify the reasoning worked)
- math
- chess
- battleship
- mastermind
- coding
- etc
in other words, testbeds or simulation environments were we can accurately calculate whether or not the reasoning was perfect. In my video this morning ( https://youtu.be/zzaEBGOVKIg ) I demonstrated that Claude seems to be able to validate and clean data. But we'll need to verify that.