EditScore/EditScore-Qwen3-VL-4B-Instruct
Adapter vs Qwen/Qwen3-VL-4B-Instruct. EditScore/EditScore-Qwen3-VL-4B-Instruct is a LoRA adapter, not full weights: rank 32, on down_proj, gate_proj, k_proj, linear_fc1, linear_fc2, o_proj, q_proj, up_proj… in 36 layers (258 adapted matrices). It modifies Qwen/Qwen3-VL-4B-Instruct at load time.