vllm-sr/mmbert-embed-32k-2d-matryoshka
Full fine-tune vs vllm-sr/mmbert-32k-yarn. Fine-tune: 100% of the weights changed, across 22 of 22 layers.
Tensors changed: Embeddings 2/2, Norms 1/44, Attention 44/44, MLP 44/44
2 derivatives checked
- vllm-sr/mmbert-embed-finance: Full fine-tune
- vllm-sr/mmbert-embed-medical: Full fine-tune
3 other repos hold byte-for-byte identical weight files.