mergekit-community/mergekit-passthrough-kfpckkg
Partial fine-tune (fewer layers) vs Undi95/Meta-Llama-3-8B-Instruct-hf. Undi95/Meta-Llama-3-8B-Instruct-hf matches part of mergekit-community/mergekit-passthrough-kfpckkg (291 of its 462 tensors; the rest of mergekit-community/mergekit-passthrough-kfpckkg has no counterpart). On the shared part: Partial change: 22% of the weights changed (attention, mlp, norms), in 8 of 32 layers.
Tensors changed: Embeddings 0/1, Norms 16/65, MLP 24/96, Attention 32/128, lm_head 0/1
1 derivatives checked
- MaziyarPanahi/mergekit-passthrough-kfpckkg-GGUF: Quantization