Additional info on intent or use case of merges

#1
by jebcarter - opened

Good morning!

I love the flood of Nemo merges coming out here - still near the top of the charts for local writing and often better (in my eyes) than the big API models - but I'd love some more information about what the intended flavor or use of some of these merges are?

At least a "hey I liked how model X did Thing Y so I merged with model Alpha whose Thing Z I dig" or sommat. :D

You are cooking strong contenders for unseating Muse with the combination of high grade prose and good instruct following but ye gads its a lot of these. :)

Good morning!

I love the flood of Nemo merges coming out here - still near the top of the charts for local writing and often better (in my eyes) than the big API models - but I'd love some more information about what the intended flavor or use of some of these merges are?

At least a "hey I liked how model X did Thing Y so I merged with model Alpha whose Thing Z I dig" or sommat. :D

You are cooking strong contenders for unseating Muse with the combination of high grade prose and good instruct following but ye gads its a lot of these. :)

Hey thanks, currently intended use section is to just fill to footer space in card. I don't want to claim anything like X model is better at Y in descriptions etc. Because everyone uses different settings and prompts this can be subjective.

If there was a more neutral way to describe models I may be into that, that disregards any personal bias.

I do sometimes test on same seed with a neutral prompt and a LLM tells me the kind of prose. I could put this somewhere on card to describe general model "flavor", I would be open to any ideas you have.

I completely understand where you are coming from - I still have flashbacks to Psyonic-Cetacean and Psyfighter's releases.

Not neutral, but as a starter what I'm usually looking at is:

What's in the training data? Adventure, story, chat?
Long or short outputs by default.
Dry or purple outputs by default.
Spatial tracking (The McGuffin is a Thing inside a drawer. The drawer is locked. I get the McGuffin -> does the model understand the drawer needs to be unlocked and opened? Just opened? Does it write about smashing the drawer and just taking it)
I know lots of people care about turn taking but that is so down to prompts/harness/text vs chat completions that I don't think there's any reliable way to go after it.

In a more narrow scope, what are you as a model cooker GENERALLY trying to work for? Chat? Story? Mix?

I try to create unique merges with good prose for roleplay/story that are also smart.

I look for the same things usually:

  • Does it output too much probably broken in some way
  • Does it output too little and not match example also bad
  • Detail retrieval
  • Instruction following
  • Prose and pacing

As for the card, I am not sure what to say.

Sign up or log in to comment