fpadovani/arb-arab-100mb-after-ppt-shuff-dyck-10mb-ckpt500_seed3407
Full fine-tune vs fpadovani/arb-arab-100mb-ppt-shuff-dyck-10mb_seed3407. Fine-tune: 100% of the weights changed, across 12 of 12 layers.
Tensors changed: Attention 48/48, Norms 25/50, MLP 48/48, Embeddings 2/2