nvidia/GR00T-H
Partial fine-tune vs nvidia/GR00T-N1.6-3B. Partial change: 80% of the weights changed (attention, embeddings, mlp, norms, other), in 27 of 27 layers.
Tensors changed: Embeddings 7/9, Attention 536/540, other 147/149, Norms 239/241, lm_head 0/1, MLP 162/166
3 other repos hold byte-for-byte identical weight files.