Test Opinion
#1
by
Duttones
- opened
I kinda of liked this model, but I found it to ignore instructions very often. It's like the model has some ideas of his own about the history.
Example:
- I have setup a scenario with instruction that no matter what, the people living in there would never awake.
- The model contanstly ignored the instruction and everyone get's awaken in about 1-2 turns.
On the other side, I've seen the model putting character's manneirisms and personality in his dialogues very often.
I wasn't able to get around that just with prompting. Unless I've put the instruction right before the bot answer. I think that makes the model kindaof biased towards something I'm unaware of.
By weight, the model is mostly Llama 3.1 8B Instruct, which I'm now suspecting may be quite opinionated about narratives.
Try adding this to the Instruct prompt see if it makes a difference? Stick to the script.