hotchpotch/bekko-embedding-v1-a25m-pt

Full fine-tune vs hotchpotch/mmBERT-L13H384-pruned. Fine-tune: 100% of the weights changed, across 13 of 13 layers.

Tensors changed: Embeddings 2/2, Norms 26/26, Attention 26/26, MLP 26/26

1 derivatives checked