Command & Parameter Hallucinations
Duplicates
- [X] I have searched the existing issues
Summary 💡
Has anybody tried to control those hallucinations when it comes to generating commands and parameters ? I don't really mind that it's sometimes substituting garbage into those arguments - but whole commands and parameters is a waste of time and API resources, especially since the agent cannot even create its own commands currently.
If was thinking of splitting up the API call to have one preparation stage for the selection of the command, and another for the parameter names - with a low temperature setting, and only use higher temperature for the actual substitutions. Should that suffice, better ideas ?
Any thoughts / ideas or local experiments ?
Examples 🌈
try shopping for carpets at midnight, it will come up with all sorts of crazy carpet related "commands" and params it wants to use
Motivation 🔦
No response
Ok, figured out a compromise I guess: One simple option would be evaluating the current action/command suggestion inside the main agent loop and check the command registry to see if the command exists, and if it supports the corresponding arguments or not.
We can then update the prompt trigger to say "ERROR: unsupported command/parameter combination" and break out of the loop by running a continue. We should probably update an error count and add an evaluation criteria that tells the LLM to reduce the number of errors.
If this is in place, we could update continuous mode to make it a little smarter/faster, because the LLM would get to see that commands are invalid before actually running them.
Thoughts ?
#https://github.com/Significant-Gravitas/Auto-GPT/issues/2987#issuecomment-1563742221
This issue has automatically been marked as stale because it has not had any activity in the last 50 days. You can unstale it by commenting or removing the label. Otherwise, this issue will be closed in 10 days.
This issue was closed automatically because it has been stale for 10 days with no activity.