| models: | |
| - model: qingy2019/Qwen2.5-Math-14B-Instruct | |
| parameters: | |
| weight: 0.35 # Strong performance in GPQA, MUSR, and MMLU-PRO | |
| density: 0.6 # Retain 60% of significant parameters | |
| - model: arcee-ai/Virtuoso-Small | |
| parameters: | |
| weight: 0.30 # Exceptional IFEval and MATH Level 5 capabilities | |
| density: 0.6 # Retain 60% of significant parameters | |
| - model: CultriX/Qwen2.5-14B-MegaMerge-pt2 | |
| parameters: | |
| weight: 0.20 # Balanced contributions to Truthful QA and MMLU | |
| density: 0.5 # Retain 50% of significant parameters | |
| - model: CultriX/SeQwence-14B | |
| parameters: | |
| weight: 0.15 # Provides diverse data and generalization | |
| density: 0.4 # Retain 40% of significant parameters | |
| - model: v000000/Qwen2.5-Lumen-14B | |
| parameters: | |
| weight: 0.10 # Enhances creative and narrative tasks | |
| density: 0.5 # Retain 50% for task diversity | |
| base_model: Qwen/Qwen2.5-14B | |
| merge_method: dare_ties | |
| parameters: | |
| normalize: true # Ensures parameter scaling compatibility | |
| int8_mask: true # Optimizes memory and computational efficiency | |
| dtype: bfloat16 | |
| tokenizer_source: Qwen/Qwen2.5-14B-Instruct | |