ReadyArt/Forgotten-Abomination-70B-v5.0
The new 70B Safeword was a great success at last! We finally worked out the Llama kinks.
So what we did this time for the Abomination merge was add some flavor from Nevoria to it.
https://huggingface.co/ReadyArt/Forgotten-Abomination-70B-v5.0
priority bumped :)
Alas, I must admit, I personally didn't have success with F-S-70B v5.0, but I am suffuciently weird :)
I very much welcome specific feedback. By didn't have success do you just mean the model wasn't for you, or were there any specific problems I could address? Also we made some settings changes vs Llamaception these worked better. I'll update the card. https://huggingface.co/sleepdeprived3/Llama-3.3-T4
These settings increase the Dry penalty and drastically change the prompt to a simpler one.
Well, I didn't say much because I don't think the model is at fault: A model that requires a specific (chat) prompt format, or SillyTavern, to work immediately doesn't work for me, because I am writing stories, and a pure assistant can't help me with that. And I absolutely need my own prompt and prompt format to work. But my specific problem was that the model generates walls of text without any formatting (such as paragraphs) - quite likely because I don't use it for chat.
I don't count this as a problem with the model - I am simply not using it as an assistant/chat model. It is, however, especially bad with my style of prompting, regardless of its other qualities.
Understandable. You may like my new Gaslight model coming out in the next couple of days. It should be less prompt reliant. I'll send you a message when it's ready.
Keep in mind that I really am an unimportant minority here. And quite the edge case - I was stuck on QuartetAnemoi for almost a year :) But sure, I will happily try it out :)