manycore-research/SpatialLM1.1-Llama-1B
Full fine-tune (fewer layers) vs meta-llama/Llama-3.2-1B-Instruct. meta-llama/Llama-3.2-1B-Instruct matches part of manycore-research/SpatialLM1.1-Llama-1B (146 of its 604 tensors; the rest of manycore-research/SpatialLM1.1-Llama-1B has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 16 of 16 layers.
Tensors changed: Embeddings 1/1, Norms 33/33, MLP 48/48, Attention 64/64