chaoyinshe/EchoVLM
Partial fine-tune (fewer layers) vs Qwen/Qwen2-VL-7B-Instruct. Qwen/Qwen2-VL-7B-Instruct matches part of chaoyinshe/EchoVLM (730 of its 1262 tensors; the rest of chaoyinshe/EchoVLM has no counterpart). On the shared part: Partial change: 79% of the weights changed (attention, mlp), in 28 of 32 layers.
Tensors changed: Embeddings 0/2, Norms 0/187, MLP 88/216, Attention 112/324, lm_head 0/1