KaLM-Embedding/KaLM-Reranker-V1-Small-R2
LoRA-style vs google/t5gemma-2-1b-1b. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, k_proj, q_proj, up_proj, v_proj changed, in 26 of 27 layers; norms, biases and embeddings are untouched.
Tensors changed: MLP 156/264, Norms 0/425, Attention 156/424, Embeddings 0/5, other 0/1