Klingspor/StarPO-4B
LoRA-style vs Klingspor/Qwen3-4B-SFT. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, k_proj, o_proj, q_proj, up_proj, v_proj changed, in 36 of 36 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/1, Norms 0/145, MLP 106/108, Attention 144/144, lm_head 0/1
2 derivatives checked
- mradermacher/StarPO-4B-GGUF: Same weights, converted
- mradermacher/StarPO-4B-i1-GGUF: Quantization