WillHeld/Delphi-25B-SimpleRL-Math

Full fine-tune vs marin-community/delphi-1e23-25Bparams-628Btokens. Fine-tune: 97% of the weights changed, across 51 of 51 layers.

Tensors changed: lm_head 1/1, Embeddings 0/1, Norms 205/205, MLP 153/153, Attention 204/204

2 derivatives checked