| models: | |
| - model: bunnycore/Best-Mix-Llama-3.1-8B | |
| parameters: | |
| density: [1, 0.7, 0.5] # density gradient, focusing on overall balance | |
| weight: 1.0 | |
| - model: USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B | |
| parameters: | |
| density: 0.6 # chemistry focus with significant contribution | |
| weight: [0.3, 0.7, 1.0] # weight gradient, increasing focus on chemistry tasks | |
| - model: Weyaxi/Einstein-v6.1-Llama3-8B | |
| parameters: | |
| density: 0.4 # focus on conversational adaptability and long-form generation | |
| weight: | |
| - filter: mlp | |
| value: 0.5 # stronger contribution to MLP tasks | |
| - filter: self_attn | |
| value: 0.7 # slightly higher weight on attention mechanisms | |
| - value: 0.5 # balanced contribution for general tasks | |
| merge_method: ties | |
| base_model: bunnycore/Best-Mix-Llama-3.1-8B | |
| parameters: | |
| normalize: true | |
| int8_mask: true | |
| dtype: float16 | |