Suggestion: increase the weights of instruction-following models.
First thank you for this already good merge.
I appreciate the idea of dropping the bagel because its inherent problem of ppl at long ctx. However, I find the resulting model does not follow instructions very well. For example, in a role playing session, I often use direct prompts (/sys in SillyTavern) instead of acting as my character to direct the story. And this model often ignores there prompts and continue as is. I presume it's a bit lacking in instruction following capability as the backbone of this merge are mostly non-instruct tuned base models/"undertrained" ones.
The bagel had a new version of 0.4 now: https://huggingface.co/jondurbin/bagel-34b-v0.4 Probably you could add it back at a lower weight (or wait for the 0.4 dpo version), or you could increase the weights of instruction-following models in the merge.
I recommend steering the story using OOC comments, seem to be working really well for this model. Alternatively, you can create a Narrator Persona and switch to them to respond. Hope this helps!
I recommend steering the story using OOC comments, seem to be working really well for this model. Alternatively, you can create a Narrator Persona and switch to them to respond. Hope this helps!
@MarinaraSpaghetti
Hi, How do you steer the story using OOC comments? Could you give me a simple example?
At the end of my last message, I simply add:
(OOC: Hey, could you please do X from now on/steer the story this way? Thank you!)
Worked when I switched present tense to past tense out of nowhere, to do some testing. And it works if you want the model to include something specific.