Finnish-NLP/Ahma-3B-Instruct
LoRA-style vs Finnish-NLP/Ahma-3B. 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 26 of 26 layers; norms, biases and embeddings are untouched.
Tensors changed: Norms 0/53, MLP 78/78, Attention 104/104
4 derivatives checked
- mradermacher/Ahma-3B-Instruct-GGUF: Same weights, converted
- ducklingcodehouse/Finnish-DentalQA-v2-merged: Full fine-tune
- ducklingcodehouse/Finnish-DentalQA-merged: Full fine-tune
- RASMUS/Ahma-3B-dpo-slerp: Full fine-tune