pinecone/ConstBERT
Full fine-tune vs google-bert/bert-base-uncased. Fine-tune: 99% of the weights changed, across 12 of 12 layers.
Tensors changed: Embeddings 2/3, Attention 96/96, other 48/50
1 other repos hold byte-for-byte identical weight files.
Full fine-tune vs google-bert/bert-base-uncased. Fine-tune: 99% of the weights changed, across 12 of 12 layers.
Tensors changed: Embeddings 2/3, Attention 96/96, other 48/50
1 other repos hold byte-for-byte identical weight files.