Pinkstack/Phi-4-mini-6b-merge
Partial fine-tune (fewer layers) vs microsoft/Phi-4-mini-instruct. microsoft/Phi-4-mini-instruct matches part of Pinkstack/Phi-4-mini-6b-merge (194 of its 338 tensors; the rest of Pinkstack/Phi-4-mini-6b-merge has no counterpart). On the shared part: Partial change: 81% of the weights changed (attention, mlp, norms), in 31 of 32 layers.
Tensors changed: Embeddings 0/1, Norms 62/65, MLP 62/64, Attention 62/64
2 derivatives checked
- mradermacher/Phi-4-mini-6b-merge-i1-GGUF: Quantization
- mradermacher/Phi-4-mini-6b-merge-GGUF: Same weights, converted