yamatazen/EerieLight-21B
Partial fine-tune (fewer layers) vs yamatazen/EtherealAurora-12B. yamatazen/EtherealAurora-12B matches part of yamatazen/EerieLight-21B (363 of its 633 tensors; the rest of yamatazen/EerieLight-21B has no counterpart). On the shared part: Partial change: 67% of the weights changed (attention, mlp, norms), in 30 of 40 layers.
Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 60/81, MLP 90/120, Attention 120/160
2 derivatives checked
- mradermacher/EerieLight-21B-i1-GGUF: Quantization
- mradermacher/EerieLight-21B-GGUF: Quantization