It's a 'frisbee' learning lora model.

frisbee = flying or on floor
frisbee, holding = in hand
frisbee in mouth = eat

recommend weight 0.8 to 1.0.

ILXL models already understand this concept, so 0.00 ILXL version of lora gives some quality boost.