Lunzima/NQLSG-Qwen2.5-14B-MegaFusion-v8.8

Full fine-tune vs Lunzima/NQLSG-Qwen2.5-14B-MegaFusion-v5. Fine-tune: 98% of the weights changed, across 48 of 48 layers.

Tensors changed: Embeddings 1/1, Norms 97/97, MLP 141/144, Attention 329/336, lm_head 1/1

1 derivatives checked