v3 vs v4
I notice a distinct difference in the writing and grammar of the v3 and v4 models. The v3 model follows well and has good grammar as compared to the v4 model I noticed a dip in proper grammar. Replies not fully punctuated and letters not capitalized, as well as the quality of the replies being less surrounding aware. Is this because of the training, or would it be the prompting? I used the same prompts for both and had completely different responses.
note: I am using LMStudio and have not tried it with other software. I could assume it would give different responses with different software as well.
Ps: v3 works amazing. It's still around fimbulvertr v2 no better or worse, just a different spice of RP.
It also feels like it is worse at creative writing (might be over fitting to pure RP).
For example, if you give v3 and v4 the same prompt.
V4 will have a more matter-of-fact type of response, so "Person A did this. Then Person A did this..."
V3 is still giving less of a matter-of-fact type of response but there is noticeable amount of more description between actions, so "Person A did this [some description of them or surrounding area]. Then Person A did this [some more details or additional actions]..."
I am not a RP person, i prefer models like the Midnight or Dark models where I could give the prompt "Tell me a story about someone watching paint dry" and it will return extremely descriptive and narratively driven type of responses.
@TheDrummer Has there been any discussion on having models more leaning away from RP and more for writing? I follow your team's BeaverAi account and their uploads but it is hard to tell what the focus of those models is.
I agree with what you're saying. In my use case It's for role-play, but I mostly test them in standalone instead of silly or other RP interfaces because I like to find out if it is sufficiently aware of its surroundings while also giving compelling replies that are not just some iteration of the thing you just prompted, but a vivid detailed response to the input.
Like I said in the earlier post, v3 works ok, but v4 is kind of flooded with misspelling and hallucinations. If it was possible to train this model by hand, or any solar based model. I believe it would work better than any out of the box models on hugging face.
But the problem lies in the fact that no one has ever made a windows based program that can train your model as you use it, like RLHF. Take into consideration how OpenAI or poe.com uses RLHF.
If there was a program that can train models as you use them, I believe it would make even a low parameter model function better than a larger one, because it is tuned to work as the user sees fit, and not the developer of the LLM.
I believe the work LLM developers do is amazing, but without being able to tune it and teach it yourself, you're locked into the training data they use.
I know you can train them with different methods like Loras or QLoras, but those are based off someone else's data and not your own. Not to mention, a lot of them are also based off response and replies from other LLMs and not user input.