Takenoko12345678/Spark-X2.5-4B-Japanese
LoRA-style vs XHToken/Spark-X2.5-4B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, g_proj, gate_proj, out_proj, q_k_v_proj, up_proj changed, in 36 of 36 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/1, Norms 0/73, MLP 108/108, Attention 108/108
3 derivatives checked
- Takenoko12345678/Spark-X2.5-4B-Japanese-GGUF: Quantization
- MC7ever/spark-darwin-5way: Full fine-tune
- Takenoko12345678/Spark-X2.5-4B-Japanese-Pitcher: Adapter