PJMixers-Dev/qwen2.5-1.5b-instruct-43-layers-zeroed
Partial fine-tune (fewer layers) vs Qwen/Qwen2.5-1.5B-Instruct. Qwen/Qwen2.5-1.5B-Instruct matches part of PJMixers-Dev/qwen2.5-1.5b-instruct-43-layers-zeroed (338 of its 518 tensors; the rest of PJMixers-Dev/qwen2.5-1.5b-instruct-43-layers-zeroed has no counterpart). On the shared part: Partial change: 53% of the weights changed (attention, mlp, norms), in 18 of 28 layers.
Tensors changed: Embeddings 0/1, Norms 34/57, MLP 52/84, Attention 120/196