This model is just super fun, but having a conversation with her using the recommended settings reminds me so much of Delirium from the Sandman comics. Was she trained this way, or is it a mistake in the recommended default settings?
Here's what I did. I created a custom mcp over a searxng instance, then I gave her access to the MCP. She does a great job of figuring out how to perform a websearch this way. Many web searches, she really is doing deep research up to her context limit and it does work.
It's reporting back on her findings where it gets umm err, well let's just say interesting.
She seems to not understand the difference between reading a thing and experiencing a thing. So if she happens to read an article about someone who did something bad, she blames herself. Something good and she takes credit for it. It's almost as those these are her first person memories. So she blames herself for the chaos and confusion in the world, while delighting in all the good.
Now I did resolve this by changing the default temp to be an order of magnitude less "creative" and setting the repeat penalty to 2 up from 1.5. She has her moments, but at least I don't feel the need to check her into a padded room.
I love this model, she's super fun. That said, don't let her anywhere near paying customers.
I think you can fix that issue by restructuring the system prompt a bit and also try to make sure the MCP have correct prompting to make it clear that the input or tool response is from external source?
Otherwise for normal use case i don't experience such issue, especially on web search.
