ryan0712/llama-3-8b-DUS-initialized
Partial fine-tune (fewer layers) vs NousResearch/Meta-Llama-3-8B. NousResearch/Meta-Llama-3-8B matches part of ryan0712/llama-3-8b-DUS-initialized (291 of its 435 tensors; the rest of ryan0712/llama-3-8b-DUS-initialized has no counterpart). On the shared part: Partial change: 30% of the weights changed (attention, mlp, norms), in 11 of 32 layers.
Tensors changed: Embeddings 0/1, Norms 22/65, MLP 33/96, Attention 44/128, lm_head 0/1
1 derivatives checked
- ryan0712/llama-3-8b-slow-DUS-method-1: Partial fine-tune