- Same data composition as Freya, applied differently, trained longer too.
- Merging with its checkpoints was also involved.
- Has a nice style, with occasional issues that can be easily fixed.
- A more stable version compared to previous runs.
My Model Settings | Feel free to use DRY or XTC or whatever meme samplers. I have zero experience with them, I can't help you there.
Training time in total was ~22 Hours on a 8xH100 Node.
Then, ~3 Hours spent merging multiple epoch checkpoints through dare_ties and model experimentation on a 2xH200 Node.
Compute spent is on my own budget and wallet, through Runpod or Vast.
Probably my last main release in a while unless things change, spent too much on this.
