Good work!
The model is great for roleplay, I like that it doesn't tend to be perverted, and at the same time it handles such scenes well. I really like that the model is consistent in the character's behavior (e.g. if a character is very moral - a nun for example, and does something immoral under the influence of alcohol, then she later returns to her previous state with remorse and a moral hangover, and does not become a perverse deviant as is often the case in other models). The model is smart and handles logic well, which is evident in the narrative. In Silly Tavern, it nicely incorporates information from character cards, the scenario, and I also noticed that it can make great use of Story String and System Prompt.
https://huggingface.co/MarinaraSpaghetti/SillyTavern-Settings/tree/main/Basic/ChatML - using this I get a really decent story, realistic and engaging.
https://huggingface.co/MarinaraSpaghetti/SillyTavern-Settings/tree/main/Customized/ChatML - and this is where something amazing begins. Of course, you can modify it to suit your needs, but even in this form MarinaraSpaghetti. roleplay takes on a new dimension.
I tested https://huggingface.co/mradermacher/Qwen2.5-Lumen-14B-GGUF/blob/main/Qwen2.5-Lumen-14B.Q6_K.gguf I still plan to see the Q8 and IQ6 versions out of curiosity.
I used: Temp: 0.7, Tok k:40, Min p:0.1, Smoothing Factor:0.25, Smoothing Curve: 1, DRY Repetition Penalty - (Multiplier0.8, Base:1.75, Allowed Length:2, Penalty Range:4096) - I know the settings are pretty strict but I don't like it when characters get 'lost' and talk nonsense. Temperature 0.9 also works great with this settings.
Great model and a nice surprise. Thank you very much.
Thank you so much! Those prompts fixed an issue I was having where it was outputting chinese in place of english words sometimes, and fixed some perspective confusion it had. So I'll definitely add those to the model card as recommendations for RP.
And I agree it definitely has a more "realistic" narrative it seems, with more depth, so that's a nice change of pace. Qwen2.5's outputs feels very different and fresh compared to Mistral and Llama so I think it has untapped potential. especially with this 14B sized model, it's the most interesting one of the new Qwen-line imo.