budget-internalization-iclr2027/nemotron3-nano-4b-1k-grpo-usefulhippo-s300

Full fine-tune vs nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16. Fine-tune: 100% of the weights changed, across 42 of 42 layers.

Tensors changed: Embeddings 1/1, other 104/147, Norms 47/64, MLP 34/34, Attention 16/16, lm_head 1/1

On the Hub · Downloads on Model Pulse