I get the right answer on the 8B model too.
It could be the quantized version failing?
My models are both 4 bit. But yeah, that could be - small models are much worse at tolerating quantization. That's why people use LoRA to recover the accuracy somewhat even if they don't need domain adaptation.
alt Hacker News
My models are both 4 bit. But yeah, that could be - small models are much worse at tolerating quantization. That's why people use LoRA to recover the accuracy somewhat even if they don't need domain adaptation.