mergekit-community/Berry-Spark-7B-Fix

Full fine-tune vs ChaoticNeutrals/Very_Berry_Qwen2_7B. Fine-tune: 97% of the weights changed, across 28 of 28 layers.

Tensors changed: lm_head 1/1, Embeddings 1/1, Norms 55/57, MLP 81/84, Attention 189/196

2 derivatives checked