Feedback on this merge compared to Pygmalion 7b
You mentioned in another discussion that you were looking for feedback.
That was four days ago, which I've learned in this field is more like four months, so hopefully this is still relevant.
I've spent probably about ~5 hours with your straight Pyg7B and about ~4 with this one.
I'm using them both through Kobold connected to SillyTavern with the extension that summarizes the chat history, which I regularly correct.
The only consistent difference I've seen is that this one is a lot more likely to spit sections of my prompt directly back at me,
especially when it is asked to act as a specific character. I've also had it use the and tags several times in responses.
Now, that happened like three times, so it could just be a weird fluke, but I figured it was worth mentioning.
I haven't noticed any definite positive differences, which makes sense cause I imagine they would be subtle, but it does feel a little more on task than the base version.
It has also proven a little more adept at professional modes of writing, short summaries, and the like.
Hope that's helpful, either way, thanks for the hard work and the models.
I see, thank you for the feedback 😄
Have your tried using this as an instruct model via vicuna's own prompting style?