This model produces strange output, and uses fancy words to avoid following any instructions without overtly refusing.

T7 merges Slimaki, Maginum Cydoms, Asmodeus v1, Asmodeus v2a, Asmodeus v2e, Magistry, and Checkpoint T6 via the della method.

It appears that della merges can become overly unnormalized. Multiple attempts to stack layers of della with normalize false and total weights around 2 eventually leads to the model producing extremely long, complex and unrelated tangents which have nothing to do with the prompt.

I am now testing ties and dare_ties to see how these compare.

ties does not allow you to merge 2501 finetunes/merges with 2503/2506/2509 the way that della does. It causes grammar collapse, forgetting words like "to".

This was broken due to using ties and also censored

Testing this now with della

Here are the full yaml configs used to create this checkpoint

Asmodeus v2e

T5

T6

T7