It was trained with NF4 weights while respecting the original Turbo model steps (unlike other trainers). Used custom scripts, slightly overbaked compared to other loras.
Tested with Haoming's A1111 Forge Neo, in bfloat16 precision.
Description
Stability was improved by calculating the loss in training while applying the lora model twice as strongly on the base model; inference happens at the recommended strength of 0.6.