AI-MO/NuminaMath-72B-CoT

Full fine-tune vs Qwen/Qwen2-72B. Fine-tune: 98% of the weights changed, across 80 of 80 layers.

Tensors changed: Embeddings 0/1, MLP 240/240, Attention 560/560, Norms 160/161, lm_head 1/1

3 derivatives checked