Thanks for mentioning the individual input models

#1
by CultriX - opened

As the title says! And congrats on the final result :)!

Thank you! Merging like this is balancing on the shoulders of many giants - and considerable trial and error. For Vimarckoso, I settled on iterative rounds of LoRAs, DELLAs, dare_ties, and SLERPs to get this.

Yours have been among the mergekit YAMLs I've studied to get this result. Good job on Broca! Are we going to give Qwen 32B a run for its money or what?

I made a Python script and am working on a small leaderboard comparison tool (see my space: tiny leaderboard. I think the heatmap espcially is really nice as well as the scrape all mergekit configurations options :) ) still a work in progress though!

At some point I want to work on della_linear as a basis for evolutionary merging. You've got great data from your dare_ties merges, and I think this will carry over!

That leaderboard is brilliant. Thank you for setting it up! I'd like to confirm my hunches about what makes our most successful models tick - and I think @jeffmeloy 's approaches with NER and minperplexity merging has been very successful on the 7B models and would likely offer a lot here, too.

(UPDATE) Surprise takeaway from the leaderboard - my impression of some merge results on the comparator led me to delete/offline them, but your leaderboard is showing where they had merit. I need to tighten up my eval process. Also, you, @sthenno , and I are very close in our results, suggesting we are using current methods well, each with our favored benchmark and features. I still think there's optimization to do to get a model ready to perform with minimal finetuning.

This comment has been hidden

Screenshot_20241231_025133.png
Hats off to everyone whose models appear here. I'm sure another round of high scores is soon to come. It's wild how close they are now! https://shorturl.at/MhmwT

Sign up or log in to comment