jsfs11

/

WestOrcaNeuralMarco-DPO-v2-DARETIES-7B-GGUF

like 1

No worries, uploaded. Please let me know what you think :)

Thank you; it works perfectly. Well, since you provided more models for testing, I also tested WildMBXMarconi and TurdusTrixBeagle while comparing them. I gave them 3 puzzles, asked to finalize the idea with the character, but also asked to create a super short sci-fi thriller and chat with them just with a few messages.
WestOrcaNeuralMarco: Replies length: Long. WestOrca feels the most engaging of all 3 models. It successfully resolved 2+ out of 3 puzzles (Replied correctly after one regeneration.) With this model, you don't need to ask, "What do you propose?" Just give it the topic, and it will give proposals and suggestions even if you aren't intending to ask for them. It's a very proactive model with long and multifaceted replies with good logical capabilities. The only downside I would say is that its proactivity may turn into a monologue, especially during creative writing. Instead of co-creating, it tends to create on its own. Without even understanding what I want to see, moreover, it may generate side characters and plot points with its start and end, which I didn't ask for, lmao. The reason for that is it begins to answer questions that it guesses could be asked. Similar behavior I saw with NeuralBeagle14, and still, I should say that here it is less pronounced, which is good.
Storytelling - Naturally, I don't expect anything extraordinary, and I would say the detective was not very logical and a bit blurred in the plot; TurdusTrix did better. We also discussed my day with it, and in casual chat, the model definitely feels better than the other two, more engaging and adaptive than TurdusTrix or WildMBX. Overall, despite some coherence issues, it compensates with creativity and engagement. I believe tactfulness can be increased with sys prompt.
Now a bit about two other models.
TurdusTrixBeagle: Replies length: Mid/Balanced. This model looked more tactful in terms of turn-based dialogue, much less proactive, engaging, or curious. It won't answer what you don't intend to ask, as WestOrca can do. Once it even said, 'I can suggest if you are interested' and waited until you gave permission. It resolved 2/3 puzzles. It also feels a bit more coherent than WestOrca in some cases but loses in the naturalness of speech that WestOrca has and adaptivity to understand mistakes, for example in puzzle. The detective story turned out to be more specific in details, better structured, and logical than that of WestOrca.
WildMBXMarconi: Replies length: Short. Isn't intended for creative writing; it was unstructured somewhat. And I wouldn't say it is more logical than WestOrca. It resolved 2/3 puzzles, still it can understnad mistake. It also replies with shorter and more concentrated replies, much like OpenChat. Just compared to WestOrca, this model summarizes its answers into shorter ones. Still, even though replies are shorter, it is more engaging than TurdusTrix.

P.S. The issue was with this puzzle: "There are 10 killers in a room. The 11th killer enters the room and closes the door behind him. He kills one of the 10 killers. How many killers remain alive in the room in total?" This was the final reply of WildMBX: "So after all these events unfold, there are still...10 - 1 + 1 = 9 Killers left in that chilling room." Like reasoning was correct but incorrect final number. At the same time WestOrca replied correctly with right reasoning so I think or this puzzle is already hidden somewhere in the WestOrca datasets, or its role-playing abilities allow it to better imagine how many people are in the room.

It was just a first impression, but I believe it is mostly correct. Thank you for your models! WestOrca feels a bit more adaptive and calm than NeuralBeagle14 and more coherent than WestLake v2. Good merge, keep it up.)

Wow, thank you for the very comprehensive feedback AdamDel, It is much appreciated and much more in-depth than anything I have been able to collate info on. I do agree with your conclusions and find the WestOrca merge to be the best (or most useable) overall for me. I will be going back to the drawing board so to speak and try and work out some combination that brings out the desired characteristics from each one. Again, thank you very much for your detailed thoughts.