Not entirely without bias
I really appreciate these models the Bloke puts out and have been testing their boundaries. What I've found is that indeed, IF pre-prompted correctly, then you can get it to do some downright offensible things.. but NOT anything. It remains strong bias against anything racial, it bends a story towards an arc of justice or making the tory come out ok in the end. If pushed to extreme violence or personal animosity, it will bias itself into a loop where it simply repeats itself.
All that to say, of course, we don't need those things or want to engage with them (obviously lol) but it does remain a hindrance to a truly unhindered bot
This is and old model in LLM lifespans and should be considered relatively obslete but not without value. I would suggest trying out some of our newer Dolphin models.
Interesting, I thought this was still fairly new. My experience with mistral, dolphin, and vicuna has been that Vicuna handles chat the best. Now I wonder though If im just using old models. Thanks for the tip
Apolgies we updated the model cards with non expiring links to our Discord and it pushes the models to recently updated. If you liked this model I would give the Dolphin 2.2 Yi 200k a try.