I see in the docs that you can assign a “weight” to your assistant messages to “skip fine-tuning” on specific messages.
Can anyone elaborate on what this does exactly? I am looking to use it to give my bot the ability to recover from its own mistakes. Is this the intended use?
For example, if you had the following convorsation:
User: What is the capital of the united states
Assistant: Los angeles, (weight=0)
User: It is?
Assistant: Sorry, I got mixed up there. The capital is Washington.
Will this work? Or will it train the gpt to answer incorrectly?