Qwen 3.5
#1
by Trilogix1 - opened
Did you bench it to see how much accuracy it losses, and do you have any plans to do the same to qwen 3.5?
I updated the model cards with MMLU benchmark: 86.2% (-1.6% from baseline). I haven't dug into 3.5 yet, but based my experience with the Qwen3 VL model, the multimodals are much harder to abliterate because they have higher order refusals. Though, I am always looking for good VL models in this size range, so maybe it's worth a shot.