Sakalti/ultiima-52B

Partial fine-tune (fewer layers) vs Sakalti/ultiima-32B. Sakalti/ultiima-32B matches part of Sakalti/ultiima-52B (771 of its 1251 tensors; the rest of Sakalti/ultiima-52B has no counterpart). On the shared part: Partial change: 71% of the weights changed (attention, mlp, norms), in 48 of 64 layers.

Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 96/129, MLP 144/192, Attention 336/448

1 derivatives checked

On the Hub · Downloads on Model Pulse