joooelw/BEHAVE-27B

Full fine-tune vs Qwen/Qwen3.8-27B. Fine-tune: 91% of the weights changed, across 64 of 64 layers.

Tensors changed: Norms 121/209, Attention 355/448, MLP 192/192, Embeddings 0/1, lm_head 0/1

1 derivatives checked