sugiv/cardvaultplus-500m
LoRA-style vs HuggingFaceTB/SmolVLM-500M-Instruct. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of q_proj, v_proj changed, in 32 of 32 layers; norms, biases and embeddings are untouched.
Tensors changed: lm_head 0/1, other 0/1, Embeddings 0/4, Norms 0/115, MLP 0/144, Attention 88/224
1 derivatives checked
- sugiv/cardvaultplus-500m-gguf: Same weights, converted (fewer layers)