mujo-labs/sandman-gemma3-1b-multitask-v2

Partial fine-tune vs unsloth/gemma-3-1b-it-unsloth-bnb-4bit. Partial change: 40% of the weights changed (attention, mlp), in 26 of 26 layers.

Tensors changed: MLP 234/384, Attention 328/536, Embeddings 0/1, Norms 0/157