AI-MO/NuminaMath-72B-TIR

Full fine-tune vs AI-MO/NuminaMath-72B-CoT. Fine-tune: 98% of the weights changed, across 80 of 80 layers.

Tensors changed: Embeddings 0/1, Norms 157/161, MLP 240/240, Attention 560/560, lm_head 1/1

2 derivatives checked