win10/nemolita-21b

Partial fine-tune (fewer layers) vs MarinaraSpaghetti/Nemomix-v1.0-12B. MarinaraSpaghetti/Nemomix-v1.0-12B matches part of win10/nemolita-21b (363 of its 633 tensors; the rest of win10/nemolita-21b has no counterpart). On the shared part: Partial change: 28% of the weights changed (attention, lm_head, mlp, norms), in 10 of 40 layers.

Tensors changed: lm_head 1/1, Embeddings 0/1, Norms 20/81, MLP 30/120, Attention 40/160

2 derivatives checked

On the Hub · Downloads on Model Pulse