Hey
Sorry to annoy you, and I'm sorry about that day I confused your model with another(I'm blind sometimes), can you test this merge I'm working on trying to merge bigger models with small ones, on that repo I merged Silicon-Maid-7B with Sheared-LLaMA-2.7B-ShareGPT, 2 trials, one at 10% and other at 33%, I have the original too(without anything on top, to compare), and I did one with Kunoichi-7B at 33% too.
Just if you have time(and if you want, it's not a obligation), I'm pretty bad at testing, and I'm not sure if the models have acquired some of the behavior of the larger models, sometimes it tends to produce gibberish unfortunately.
Okay, so I tested Silicon 33% first and it is not what I would consider coherent. It is unable to speak in complete sentences and has trouble keeping up with the conversation. Then I tested Kunoichi and it is unable to respond in a manner that makes sense during riddles, but it seemed to handle conversation and roleplay relatively well. Silicon 10% was the best of the the mixed models and responded quite well to both the riddle and roleplay (though it did not correctly answer the riddle, though that is rare in 3B models.) Finally I tested the original to see where baseline was and saw quite a bit of it in the others. It sometimes spoke in sentence fragments like Silicon 33%, but otherwise performed well. I would still say that Silicon 10% was the best of the batch.
Thank you so much for trying, here I did a bigger merge it seems kinda odd too. On Kunoichi 33% merge is supposed to be kinda smoother/different than the others, but idk I'm throwing things at the wall without any ideia if it's sticking/working or not. I agree with you I liked too the 10% silicon after running a bit of rp with it(I thought it was a placebo, but now it doesn't looks like that).
Here my branch where I'm trying to improve this weird merge.
And if you have any suggestions of models or anything please contact me.
Thanks again for the help 🤗