nvidia/EGM-4B-SFT
Full fine-tune vs Qwen/Qwen3-VL-4B-Thinking. Fine-tune: 91% of the weights changed, across 36 of 36 layers.
Tensors changed: Embeddings 1/4, Norms 129/249, MLP 112/220, Attention 144/240
Full fine-tune vs Qwen/Qwen3-VL-4B-Thinking. Fine-tune: 91% of the weights changed, across 36 of 36 layers.
Tensors changed: Embeddings 1/4, Norms 129/249, MLP 112/220, Attention 144/240