sudden dune
Been redoing my Mio model for a bit on Ov2. Have like 40 mins for a dataset which I know is crazy high but it is what it is. I've got it to like 450 epoch/45K steps before overtraining but idk, it sounds a bit off. Like the esses and breaths sound choppier than the model I did on the OG pretrains. Some weird humming and chopping artifacts too.
Does Ov2 encourage smaller datasets than that or is this just something dodgy in my dataset? Or perhaps a mix of both. Thanks!
tulip oxide
Been redoing my Mio model for a bit on Ov2. Have like 40 mins for a dataset whic...
OV2 does encourage smaller datasets, I believe
And it has some problems with artifacting n stuff. I'd just recommend using the normal ones unless you have small datasets (less than 10 mins)
But again, I haven't tested Ov2 enough, so take that with a grain of salt
bronze valve
i would say the same thing based on my testing
i would even say less than 5min maybe
foggy zealot
i try on 2 min with same epoch so its have artifacting
gleaming berry
from my testing using large datasets will generally just sound the same when you train it for long enough for hours using 0v2 and V2
but on smaller datasets its better to use 0v2, for longer it doesnt really matter from my test
gaunt ferry
Been redoing my Mio model for a bit on Ov2. Have like 40 mins for a dataset whic...
no it works fine with large dataset
30 epochs and 2h of data
tulip oxide
no it works fine with large dataset
but eh
if you're on a long dataset you might as well use normal pretrains
gaunt ferry
im just testing new pretrains
bronze valve
its not making it worse or anything, i just always liked the output more without the new pretrain. except on short datasets
sudden dune
I guess it's a mix of both perhaps lmao. I hear some of the same artifacting in the OG Mio model so yeah