| merge_method: sce | |
| base_model: Qwen/Qwen2.5-32B # Pivot model (zero weight) | |
| dtype: bfloat16 | |
| parameters: | |
| select_topk: 1.0 | |
| models: | |
| # Pivot model (explicitly zero-weighted) | |
| - model: Qwen/Qwen2.5-32B | |
| parameters: | |
| weight: 0.0 # Zero contribution to merged weights | |
| # sparsity: 0.0 # Optional: enforce sparsity if needed | |
| # Target models with assigned weights | |
| - model: Qwen/QwQ-32B | |
| parameters: | |
| weight: 0.95 # Dominant contributor | |
| - model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B | |
| parameters: | |
| weight: 0.01 # Minimal contribution | |
| - model: NovaSky-AI/Sky-T1-32B-Flash | |
| parameters: | |
| weight: 0.04 # Minimal contribution | |