A clever model
Very clever model, handles logic nicely. This can be seen even when creating stories. Great for roleplay, with the settings recommended by the creator it is incredibly consistent, it does not confuse characters or plots.
Out of curiosity, I gave him my set of puzzles, and I was very surprised. When solving step by step, with 10 attempts, 2-3 solves all the puzzles (this is a record), and in the rest of the cases, almost all of them are consistently solved.
There is a noticeable improvement in its capabilities compared to version 8b. Well done!
You should really share your praise with the model creator (too) :)
I did.
Oi, naughty me simply assumed you didn't without checking!
You intrigued me, and I tried it out. I can understand the allure, the model is quite open, verbose and has good writing. Actually pretty good if you don't want rails. But if you want it to follow instructions or complex relationships it fails quite badly (mostly simply by ignoring instructions). Still, pretty amazing, and presumably a great chat partner.
as a sidenote, i usually don't go lower than 70b models, so I have no good comparison to smaller models. It does hold its own against some much bigger models though. As for recommendations, I am probably the wrong person to ask, but recently, v000000 revived the nearswap algorithm (the ones with t0.0001 in the name) and has seen some success (I only had time to test https://huggingface.co/v000000/TripletBoreas-7B-t0.0001 so far, which was far from perfect, but it was very good at instruction following. I am doubtful it works with llama-3 though, but time and experiments will tell).