Open-Assistant
Open-Assistant copied to clipboard
LAION-AI
Accepting critique dataset
We would like to create a dialog dataset where the bot answers some question, and the user corrects that answer, and the bot acknowledges the correction and makes the correction.
Example 1.
User: Write me a program to do X
Rosey:
the bot generates software code (we use bugger to inject an error -cc @furlat).
User: There appears to be a mistake with the comparison of X <= Y.
Rosey:
You are right. the comparison should by x==y because ..... Here is the corrected code.
the code the bot would generated if we didn't use bugger.
Example 2. Data mining wikipdia for instructions. User: What is anarchism? <-- main topic Rosey: Anarchism is a political philosophy and movement that is acceptance [we do an antonym replacement of the word "skepticism"] of all justifications for authority... User: Are you sure it's about acceptance of all authority? Rosey: You are right. It is the 'skepticism' of all authority. Here's the corrected answer: The actual wikipedia text
Example 3: Data mine math help websites or inject math errors directly into equations. User: Solve me a problem X. Rosey: Sure, the answer is Y_1. User: I don' think that's right. I think the answer is Z. Rosey: You are right. I was reading the problem as X_1 (injeceted error). Here's the correct answer: Y.
It seems a little bit closer to the other issue I'm in charge of. I could start this right after the other.
Could you also include dialogs where the user critique is incorrect and the bot does not make a correction
