Cursed?
How cursed is this model?
I can already imagine it doing messed up things, getting all philosophical, then horny.
Waiting for quants.
It's a weird model to say the least. I'm going to mess around with it further.
It performs okay with glade, it is rather descriptive. Which I guess would be the instruct model.
And the evil kinda shines through into characters. Straight to physical violence :3
And yeah it can get into it quick 😭
Nevermind, he went from into it to violence again. (Win 🐥)
Unsure how it will behave in quantized formats though.
And I haven't a clue how to use llamacpp with all the bpe stuff. I've tried and failed with the quantization script 😶🌫️
I've asked for quants here, they estimate a day or so.
https://huggingface.co/mradermacher/model_requests/discussions/47
How cursed is this model?
I can already imagine it doing messed up things, getting all philosophical, then horny.
Waiting for quants.
Update - there's already a Q8 here.
mradermacher/Merge-Mayhem-L3-V2-GGUF
2nd update - all the quants now, yippee
It's not as cursed as I thought, it looks stable.
It's not as cursed as I thought, it looks stable.
It does seem to behave, better yet with a custom rope scaling. I probably should've change the config to the original llama3. A rope scale of 1638400 performs much better than the 4000000 at 16K