Slerp Merging

#1
by deleted - opened

It's incredible that you can merge 4 LLMs together yet the outputs remain coherent. I'm assuming that has something to do with how much better spherical linear interpolation is compared to weight averaging.

This got me wondering if any combination of Mistrals can be merged? Are there compatibility issues (e.g. tokens)? Do you need to get permission first? Just wondering because the "smartest" Mistral I've come across is Dolphin 2.1, while the one that produced the most human-like responses is Starling alpha. Is there a reason these two couldn't be merged?

Edit: I guess this has kinda been done. Was looking around and OpenHermes2.5 is very similar to Dolphin 2.1, and neural chat is similar to Starling. https://huggingface.co/Weyaxi/OpenHermes-2.5-neural-chat-v3-3-Slerp

It's incredible that you can merge 4 LLMs together yet the outputs remain coherent. I'm assuming that has something to do with how much better spherical linear interpolation is compared to weight averaging.

This got me wondering if any combination of Mistrals can be merged? Are there compatibility issues (e.g. tokens)? Do you need to get permission first? Just wondering because the "smartest" Mistral I've come across is Dolphin 2.1, while the one that produced the most human-like responses is Starling alpha. Is there a reason these two couldn't be merged?

Edit: I guess this has kinda been done. Was looking around and OpenHermes2.5 is very similar to Dolphin 2.1, and neural chat is similar to Starling. https://huggingface.co/Weyaxi/OpenHermes-2.5-neural-chat-v3-3-Slerp

Good idea. I'm gonna this.

Sign up or log in to comment