kechengcode/Qwen2-5B-16Layers
Partial fine-tune (fewer layers) vs Qwen/Qwen2-7B. kechengcode/Qwen2-5B-16Layers matches part of Qwen/Qwen2-7B (195 of its 339 tensors; the rest of Qwen/Qwen2-7B has no counterpart). On the shared part: Partial change: 29% of the weights changed (attention, mlp, norms), in 6 of 16 layers.
Tensors changed: Embeddings 0/1, Norms 12/33, MLP 18/48, Attention 42/112, lm_head 0/1