latishab/turnsense
LoRA-style vs HuggingFaceTB/SmolLM2-135M-Instruct. 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 30 of 30 layers; norms, biases and embeddings are untouched.
Tensors changed: Norms 0/61, MLP 0/90, Attention 120/120