ItsnotAilabs/Auro-110M
Full fine-tune (fewer layers) vs EleutherAI/pythia-160m-deduped. ItsnotAilabs/Auro-110M matches part of EleutherAI/pythia-160m-deduped (51 of its 184 tensors; the rest of EleutherAI/pythia-160m-deduped has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 4 of 4 layers.
Tensors changed: Embeddings 1/1, Norms 18/18, Attention 16/16, MLP 16/16