General impression
I didn't notice much of a difference with the old version, BUT the model has obvious problems understanding the environment. I did a small test: I left my character in the kitchen and asked the bot to go into my room, take a towel from the closet and take it to the bathroom. It would seem like a simple task, but for some reason it decided that I was in my room and started knocking on the door, although my character was downstairs in the kitchen.
A small wish, the model lacks emotions and brightness, it is very monotonous in their expression. There are no outbursts of anger or, conversely, very bright positive emotions. I would like the model to cling to some character traits, visual and emotional, emphasizing them.
An attempt to add hints to the model leads to the fact that it begins to understand the context worse and these hints only hinder it.
Your feedback is duly noted, and I agree completely. I need to step up my game for the next release!
Your feedback is duly noted, and I agree completely. I need to step up my game for the next release!
This may be a silly question, but why don't you use the same dataset that was used to create MythoMax? It has proven to be effective...
Your feedback is duly noted, and I agree completely. I need to step up my game for the next release!
This may be a silly question, but why don't you use the same dataset that was used to create MythoMax? It has proven to be effective...
MythoMax was a merge, and thus there are no datasets to be trained on. Even if there were, we're talking about an era where GPT 4 only just had been released and a lot of folks where still used to GPT 3.5 quality writing. ;-)
Your feedback is duly noted, and I agree completely. I need to step up my game for the next release!
This may be a silly question, but why don't you use the same dataset that was used to create MythoMax? It has proven to be effective...
MythoMax was a merge, and thus there are no datasets to be trained on. Even if there were, we're talking about an era where GPT 4 only just had been released and a lot of folks where still used to GPT 3.5 quality writing. ;-)
Okay, I didn't know that before. Have you thought about joining forces with other model authors? For example, with Undi95 or Sao10K, I think it would be a masterpiece model, as you and these authors have your own features and twists that could complement each other.
I like the idea of the personality training and how you've set everything up, but the model has a strong and prevalent positivity bias that will warp time and space to make sure the player character is helped and unharmed unless specifically told how not to.