abideen/MegaQwen-120B
Partial fine-tune (fewer layers) vs Qwen/Qwen1.5-72B. Qwen/Qwen1.5-72B matches part of abideen/MegaQwen-120B (963 of its 1683 tensors; the rest of abideen/MegaQwen-120B has no counterpart). On the shared part: Partial change: 72% of the weights changed (attention, mlp, norms), in 60 of 80 layers.
Tensors changed: Embeddings 0/1, MLP 180/240, Attention 420/560, Norms 120/161, lm_head 0/1
1 derivatives checked
- mradermacher/MegaQwen-120B-GGUF: Quantization