Reliable and stable RP oriented model
Thank you for your model!
- PROS
Understanding of multiple characters card.
Understanding of complex instructions (Chatbot companion with RP abilities)
Attentive to user actions; doesn't push scenario forward jumping from one scene to another.
Reply length. From my tests with quotes it doesn't exceed 350 tokens and mostly stays within the range of 200-250 with a few paragraphs and around 110-170 tokens with asterisks.
Most of the time there are no issues with markdowns.
Seemingly balances reply length between non-verbal actions and direct speech with asterisks; with quotes I was less fortunate.
It is able to use both he/she and you/I pronouns in non-verbal actions (tested with asterisks)
It replied correctly to all 3 puzzles I gave it. Impressive indeed; quite rarely quality for 7b models. I want to add a screenshot.
In the role of a chatbot I find it more dialogue and less instruction/task oriented. So 4.5.3 might be good for dialogue or creative brainstorming but not for tasks. - CONS
It may misunderstand nuances of speech, sounds slightly less vivid (image bellow) and toxic than 4.5. Not that it matters much, since 4.5 had its own quirks.(unbalanced length using quotes, moving the scene too quickly, and it was less coherent if I'm allowed to compare Q4 and Q8)
Replies are different from ScientificWitchery-Beta Rrrrr... it is personal xD
To summarize, it is the most stable 7B RP-oriented model I have tried so far. It is coherent, proactive but attentive, with mid replies length. Unlike previous 7b models, the new one is stable, and that's what matters the most. The puzzle reveals differences in the model's speech style. And this can be seen in all RP.
(Left-Side)4.5.3; (Middle)4.5; (Right-Side)0.7 is simply lost.
Thank you! I'll probably have to wait a bit to recreate ScientificWitchery, it consists of 21 models, and it's done in 4 different merges, 2 LogicalEngines and 2 CreativityEngines. I wasnt sure if it was any good as it seems to have issues from testing, but I'll have a look at it again.
Thanks for the feedback as always! 💛
Uff...ehh...21 you say? A complex work, like Megamind huh? Unfortunately, I wasn't trained with LogicalEngines to grasp it in time! Maybe you noticed something I didn't, because from my testing, I saw coherence issues with Lite, and Beta was the most coherent model (same as 4.5.3) I've been using lately. It wasn't as NSFW-oriented as 4.5 was, and no need, it had a lot of pros and it is different from this one. I appreciate your work in any case, and of course not pressuring you... no no no no no of course not!
I am happy with LemonadeRP 4.5.3 on an ongoing basis, though. Because the last time, and I don't mean only your models, I didn't find any stable RP oriented 7b models to use, unlike this one. Most of them have their own pros, but are unstable in the long run. You know, I only now noticed that the image you used for this model fully reflects its behavior; if that makes sense. 💙