Wow
Just wanted to express my appreciation for this model. This was the surprise of the month for me. I try out a few models on occasion, for uncensored stories and flexible experiences. It's important to me that the model responds to me, follows me creatively and tells the story with suitable ideas. Hardly any model can do all this at once. I didn't really expect anything from this particular model, as it didn't seem to point in my direction of interest. It was only the story aspect that made me curious. But my first experiences with Q5 and now Q6 (i1) have been amazing. It makes a rock-solid impression, with pleasant language. It doesn't ramble on pointlessly and still decorates everything very nicely. It responds very sensitively and “thoughtfully” to cues, which feels great, as if it really understands what it's talking about instead of just randomly generating appropriate sentences. Of course it's still an illusion, but a pleasant one in this case. Of course it makes mistakes in terms of content, I didn't expect anything else from a 12B, but even here much less than with my last tested models up to 20B. Currently it's my favorite here on Huggingface. Thanks for that!
Interesting 🤔
Hi! Thanks for the feedback! I'm glad to hear that I was successful in boosting the creativity and storytelling flow. May I ask how you use this model? Do you do instruct-guided storywriting in Alpaca?
Hi theDrummer.
I am using it with SillyTavern, using a modified sequence setting for DreamGen. Don't ask me how I ended up with these exotic settings, but they work very good with all llama-3 based models. And beside of the initial instructions in the system prompt, I barely use instructions during the story. I try to achieve a role play-like adventure feeling from the first person perspective of my main protagonist. But I found out that llama-3 models, and Rocinante in particular, seem to prefer writing novels instead of chatting. So I have it act as a novel writer and just add my parts into it in character, whenever I feel like. This works like a charm, way better then actual role playing does, and the model picks up every cue that I throw at it, with amazing depth of understanding. I just scratch the surface and it dives in deep. It doesn't work as well with instruction based story writing, which is odd enough. Sometimes it follows them good, like between two scenarios, and at other times it just ignores them completely. Only downside of the model to me is, that it tends to be way too optimistic, aiming at a 'happily ever after' and harmony very quickly, where I would prefer more struggling and even feelings of despair or depression. But this is a common tendency among today's models, with the exception of the very interesting L3-Umbral-Mind-RP line of models, which unfortunately isn't even half as good in creative story telling as Rocinante is.
Fun story: Rocinante recently ended a story with something like: "If you enjoyed my story, please leave a comment below."
That's interesting with this DreamGen setting. Could you share the specific settings you use with this model?
https://huggingface.co/dreamgen/opus-v1.2-llama-3-8b/tree/main/configs/silly_tavern/v1
https://huggingface.co/dreamgen/opus-v1.2-llama-3-8b/tree/main/configs/silly_tavern/v2
I found something like this... and the effect is actually very interesting - the story takes into account more the information contained in the cards and the scenario. Did you have these settings in mind?
Hey Danioken,
yes, those look very similar to the settings I started with, and was very pleased with the results. They run so smoothly with L3 models. From there I started to mess around with the settings, to mold them more to my liking. Sometimes a single words seems to determine if it gets worse or better, and I didn't figure out what makes it better or worse. I just uploaded my current settings, still experimental, hope this did work:
https://huggingface.co/datasets/Huegli/Sillitavern_L3_Settings/tree/main
Note: my instruct setting is build as a base, it won't do any good on it's own! really, don't even try it, the results are broken. You always need to add a link to a perspective as a world description, and give the story a very good opening dialogue, to get things running and to define the mood. I built it this way to have one instruction set for all kinds of stories, and to link the story's perspective directly to the story, so I would not have to think about it later, when I have character cards or stories with different perspectives..
The world - perspective - char - 1st person.json is a very interesting experiment, meant to run rocinante as a story writer instead of a chat model. Works amazingly. This preset would write the story from the first person perspective of {{char}}, leaving {{user}} as a reader. Rocinante is very powerful and creative, so enjoy it writing the story by itself. Kept me up very late for several days, because I wanted to know how the story continues. If you want to add your own ideas during the story, just edit it's last reply accordingly.
world - perspective - user - char - 3rd person.json would be the classical roleplay approach, with {{user}} roleplaying herself/himself and {{char}} being the partner in the roleplay, described in 3rd person view.
You could build every perspective this way, just add another world info for the desired perspective.
Always remember to have a meaningful opening dialog that defines the style, the mood and the typeface. Then you won't need any examples.
Hey Danioken,
yes, those look very similar to the settings I started with, and was very pleased with the results. They run so smoothly with L3 models. From there I started to mess around with the settings, to mold them more to my liking. Sometimes a single words seems to determine if it gets worse or better, and I didn't figure out what makes it better or worse. I just uploaded my current settings, still experimental, hope this did work:
https://huggingface.co/datasets/Huegli/Sillitavern_L3_Settings/tree/main
Note: my instruct setting is build as a base, it won't do any good on it's own! really, don't even try it, the results are broken. You always need to add a link to a perspective as a world description, and give the story a very good opening dialogue, to get things running and to define the mood. I built it this way to have one instruction set for all kinds of stories, and to link the story's perspective directly to the story, so I would not have to think about it later, when I have character cards or stories with different perspectives..
The world - perspective - char - 1st person.json is a very interesting experiment, meant to run rocinante as a story writer instead of a chat model. Works amazingly. This preset would write the story from the first person perspective of {{char}}, leaving {{user}} as a reader. Rocinante is very powerful and creative, so enjoy it writing the story by itself. Kept me up very late for several days, because I wanted to know how the story continues. If you want to add your own ideas during the story, just edit it's last reply accordingly.
world - perspective - user - char - 3rd person.json would be the classical roleplay approach, with {{user}} roleplaying herself/himself and {{char}} being the partner in the roleplay, described in 3rd person view.
You could build every perspective this way, just add another world info for the desired perspective.
Always remember to have a meaningful opening dialog that defines the style, the mood and the typeface. Then you won't need any examples.
Thank you. I'm just starting to play Silly tavern and I've just discovered word info (lora books) :-) "world - perspective - user - char - 3rd " - very interesting. Do you know any interesting model for RP llama 3.1?
Rocinante is actually the best model up to 20B I ever tried. It sometimes can be a bit stubborn at the start, because it doesn't really like instructions. But as I already wrote: It loves story telling, so just give it a good start, and it will be rolling. I might even remove some instructions from the system prompt, because the model seems to run smoother, the less instructions it actually gets. If you don't like my method with the linked perspectives, you could as well just copy the text and put it at the end of the system prompt. My setup is designed in a way, that this is exactly what it does anyway, when you link a perspective.