SCLM Option B - Deep Integration
Architecture
- State dimension: 384
- Injection layers: [4, 8, 12, 16, 20, 24]
- EARCP params: 94.6M (5.25% overhead)
- Experts: 3
Features vs Option A
- Deeper integration with attention and FFN
- More injection points (6 layers)
- Larger state dimension (384)
- Multi-head attention pooling for encapsulation
Author
Mike Amega (Ame Web Studio)
Model tree for amewebstudio/sclm-modelEarcp-optionB
Base model
meta-llama/Llama-3.2-3B-Instruct