So, slowly, but surely, I work on remaking Hiveminds from scratch. Once again.
I've learned many things. Tried a lot of different stuff, too.
I've tried Prodigy for my adaptive learning rate. It worked really well, don't get me wrong, but perhaps I need for it to work... less well, so to speak. So I turned my mind back to AdamW8Bit. with manual searching for learning rates.
I've tried to teach multiple Hiveminds, only to find out that they don't want to mix if datasets are different and the batch size set too high. So I've remade datasets entirely.
I've found out that using some keywords do fix some style, but at the hefty price - by making it less original. So I've redoing my datasets using only authentic keywords, without adding masterclass tags. Although, with AI, you can't be sure in anything.
Does make me depressed a bit.
| Viewed: |
12 times |
| Added: |
1 month, 3 weeks ago
27 Jan 2025 20:08 CET
|
|
|
