dphn/dolphin-vision-72b
Partial fine-tune (fewer layers) vs Qwen/Qwen2-72B. Qwen/Qwen2-72B matches part of dphn/dolphin-vision-72b (963 of its 1415 tensors; the rest of dphn/dolphin-vision-72b has no counterpart). On the shared part: Partial change: 47% of the weights changed (attention, embeddings, lm_head, mlp, norms), in 79 of 80 layers.
Tensors changed: Embeddings 1/1, MLP 108/240, Attention 252/560, Norms 1/161, lm_head 1/1