TIGER-Lab/Mantis-8B-siglip-llama3
Partial fine-tune vs TIGER-Lab/Mantis-8B-siglip-llama3-pretraind. Partial change: 89% of the weights changed (attention, lm_head, mlp, norms, other), in 32 of 32 layers.
Tensors changed: Embeddings 0/4, Norms 65/177, MLP 96/208, Attention 128/348, other 4/5, lm_head 1/1
1 other repos hold byte-for-byte identical weight files.