This is an amazing model! 👏
I don't know how carefully you tested/weighted your choices for this merge, if it's a fluke or genius. I just wanted to shout in the aether how much I appreciated your model.
Also, I'd like to know if you ran your sampling with Top NSigma, DRY or mirostat. It's quite resilient, but I'd like to test it further from the merger's perspective, if possible.
I don't know how carefully you tested/weighted your choices for this merge, if it's a fluke or genius. I just wanted to shout in the aether how much I appreciated your model.
Also, I'd like to know if you ran your sampling with Top NSigma, DRY or mirostat. It's quite resilient, but I'd like to test it further from the merger's perspective, if possible.
Thanks! Yeah, I did the weight intentional, and tested for couple of days. Had like 4 other merge settings that failed. My samplings are pretty casual, just temperature 1, min_p around 0.05-0.1, top_k adjust 85-95.
Had like 4 other merge settings that failed.
Hopefully a few gave you good insights or ideas for future ones. I had one turn full Naked Lunch on me with very tame samplings setting, failures can be interesting sometimes.
My samplings are pretty casual, just temperature 1, min_p around 0.05-0.1, top_k adjust 85-95.
Thank you! I'll give it a try. If it matters to you, I tested it with Temp 0.85, Min-P 0.1, Rep-Pen 1.05, Top-P 1, Top-K 0, DRY 0.8.