AI45Research/SafeWork-RM-Value-72B

Full fine-tune vs Qwen/Qwen2.5-VL-72B-Instruct. Fine-tune: 97% of the weights changed, across 80 of 80 layers.

Tensors changed: Embeddings 1/2, Norms 123/226, Attention 687/688, MLP 436/436, lm_head 0/1

2 derivatives checked