Diamond-Arconte-13B
Models used:
Finally, something that has common sense! If there was something I didn't like about my models, it was that most of them suffer from logical errors. For example, you have the ring of power from LOTR, and a character may claim that he has the ring instead, or your brother might claim at some point to be your father instead. Those types of errors could be because of bad merging, spelling mistakes, bad prompting, or simply derived from the quantified nature of the model; Diamond-Arconte still has those types of errors, but it gets one thing wrong at a time instead of five in one go, and usually a swipe is all you need to solve the issue.
Diamond-Arconte also has an implicit understanding of contextual tropes. Dragons? Caves or dungeons. Mad scientist? luxurious mansions and evil lairs. Student council president? Japanese high school. You don't need to clarify some of the contextual information about characters because of the model's broad understanding of different topics/tropes. It can also interpret vague personality traits, like "conflictive" or "easygoing," without any need to explain how they manifest themselves in the character; you would still get better results if you explained them yourself. This "generic" understanding of tropes might or might not be what you are looking for, but give it a try.
Due to the DiamondForce inheritance, when you use the default Alpaca instruct you might get some messages wrapped in [] from the model, like the model explaining what happened or what is about to happen, sometimes it gives unkown information about a secondary character (which is not present anywhere in the context), and so on. I even once got an instance of "self-awareness" when I triggered a lorebook entry: [Note: This information about the (lorebook entry) was not present earlier in the conversation. However, continuing with this information now doesn't break the story or create any confusion, as it allows for further story progression.]
This [] messages usually manifests if you mix the Alpaca instruct with [INST]-[/INST] of DiamondForce, like using [INST]###Instruction:{uterance}[/INST]
While I haven't tested DiamondForce, I am guessing it has some censorship in it. I only got one case of censoring in which a [] message got appended at the beginning of the model's response, only for it to ignore its own warning and answer like nothing happened. Regardless, it could still be censored in other ways that I am not aware of.
One negative aspect of this model (or positive, depending on who you ask) is that it is token dispenser; I would argue that the average length response is around 170330 tokens, and it will in most cases reach 350450 tokens if you let it cook. This is probably due to the Limarp inheritance in Emerhyst.
One weird quirk the model has is that it will end responses with a single '. To my understanding, removing the ' or leaving it there doesn't change future responses at all.
In conclusion:
-
Good contextual grasp
-
Excelent grasp on characters personalities
-
Logical/grounded/"predictable" responses
-
Good for 1-on-1 chatbots
+- Long responses (possibly my samplers/instruct fault)
+- Overlord/"self-aware" AI messages occasionally (might be funny)
-
Unknown level of censorship
-
Long responses sometimes loop on a certain topic, which can be annoying
-
Weak story-telling/narrator capabilities (I am not sure about this, but I am certain that at best they are average)
Regarding samplers, the model works well with high temperatures (even as high as 4). I like to use neutral samplers except for temperature between 0.56 to 1.08 and Top A between 0.09 to 0.17, along with a penalty range of 640 and a rep penalty between 1.03 to 1.07. A higher rep penalty stimulates the model's creativity, but anywhere past 1.1 and it will start to be incoherent. If you feel the model is being bland, raise the temperature to something like 1.8 and the rep penalty.