neulab/Pangea-7B
Full fine-tune (fewer layers) vs Qwen/Qwen2-7B-Instruct. Qwen/Qwen2-7B-Instruct matches part of neulab/Pangea-7B (339 of its 735 tensors; the rest of neulab/Pangea-7B has no counterpart). On the shared part: Fine-tune: 93% of the weights changed, across 28 of 28 layers.
Tensors changed: Embeddings 0/1, Norms 57/57, MLP 84/84, Attention 196/196, lm_head 1/1
1 derivatives checked
- neulab/CulturalPangea-7B: Partial fine-tune