Varho/Marco-Mini-Global-TV-Instruct

Full fine-tune vs ATH-MaaS/Marco-Mini-Base. Fine-tune: 100% of the weights changed, across 28 of 28 layers.

Tensors changed: Embeddings 1/1, Norms 111/113, MLP 2773/2773, Attention 112/112, lm_head 1/1

1 derivatives checked