josephmayo/Holo-3.1-4B-Coder
LoRA-style (fewer layers) vs Hcompany/Holo-3.1-4B. josephmayo/Holo-3.1-4B-Coder matches part of Hcompany/Holo-3.1-4B (426 of its 724 tensors; the rest of Hcompany/Holo-3.1-4B has no counterpart). On the shared part: Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of k_proj, o_proj, q_proj, v_proj changed, in 8 of 32 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/1, MLP 0/96, Norms 0/105, Attention 32/224
4 derivatives checked
- mradermacher/Holo-3.1-4B-Coder-i1-GGUF: Quantization
- mradermacher/Holo-3.1-4B-Coder-GGUF: Same weights, converted
- josephmayo/Holo-3.1-4B-Coder-GGUF: Same weights, converted
- usermma/Holo-3.1-4B-Coder-abliterated-failspy: Partial fine-tune