CodeGoat24/UnifiedReward-2.0-qwen35-27b

Full fine-tune vs Qwen/Qwen3.5-27B. Fine-tune: 95% of the weights changed, across 64 of 64 layers.

Tensors changed: lm_head 1/1, Embeddings 3/4, MLP 302/304, Attention 499/556, Norms 242/319

3 derivatives checked