LiAutoAD/Ristretto-3B
Full fine-tune (fewer layers) vs Qwen/Qwen2.5-3B-Instruct. Qwen/Qwen2.5-3B-Instruct matches part of LiAutoAD/Ristretto-3B (433 of its 891 tensors; the rest of LiAutoAD/Ristretto-3B has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 36 of 36 layers.
Tensors changed: Norms 73/73, MLP 108/108, Attention 252/252