paudelanil/trocr-devanagari-2
Full fine-tune (fewer layers) vs amitness/roberta-base-ne. amitness/roberta-base-ne matches part of paudelanil/trocr-devanagari-2 (106 of its 366 tensors; the rest of paudelanil/trocr-devanagari-2 has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 6 of 6 layers.
Tensors changed: lm_head 5/5, Embeddings 4/5, Norms 24/24, Attention 48/48, other 24/24