powermove72/SoMix2-xb
Partial fine-tune (fewer layers) vs MaziyarPanahi/TheTop-5x7B-Instruct-S3-v0.1. MaziyarPanahi/TheTop-5x7B-Instruct-S3-v0.1 matches part of powermove72/SoMix2-xb (291 of its 453 tensors; the rest of powermove72/SoMix2-xb has no counterpart). On the shared part: Partial change: 20% of the weights changed (attention, lm_head, mlp, norms), in 6 of 32 layers.
Tensors changed: MLP 18/96, Norms 13/65, Attention 24/128, lm_head 1/1, Embeddings 0/1
1 derivatives checked
- mradermacher/SoMix2-xb-GGUF: Quantization