kurogane/sarashina2.2-5.48B-instruct-no-tuning
Partial fine-tune (fewer layers) vs sbintuitions/sarashina2.2-3b-instruct-v0.1. sbintuitions/sarashina2.2-3b-instruct-v0.1 matches part of kurogane/sarashina2.2-5.48B-instruct-no-tuning (291 of its 507 tensors; the rest of kurogane/sarashina2.2-5.48B-instruct-no-tuning has no counterpart). On the shared part: Partial change: 63% of the weights changed (attention, mlp, norms), in 24 of 32 layers.
Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 48/65, MLP 72/96, Attention 96/128
1 derivatives checked
- mradermacher/sarashina2.2-5.48B-instruct-no-tuning-GGUF: Same weights, converted