stockmark/Stockmark-Nemotron-3-Nano-Omni-JapanDocReader
Full fine-tune vs nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16. Fine-tune: 91% of the weights changed, across 52 of 52 layers.
Tensors changed: Norms 144/567, Embeddings 2/4, other 148/296, MLP 1715/1812, Attention 152/320, lm_head 1/1