| models: | |
| - model: VAGOsolutions/SauerkrautLM-v2-14b-DPO | |
| parameters: | |
| weight: 0.25 # Prioritize top IFEval | |
| density: 0.6 # Keep a large portion for strong factual baseline | |
| - model: allknowingroger/QwenSlerp6-14B | |
| parameters: | |
| weight: 0.25 # High weight for MATH and balanced reasoning | |
| density: 0.6 # Retain robust reasoning capabilities | |
| - model: CultriX/SeQwence-14B-EvolMerge | |
| parameters: | |
| weight: 0.20 # Important for best BBH and near-top MuSR | |
| density: 0.5 # Moderate density to ensure these strengths blend well | |
| - model: CultriX/Qwen2.5-14B-Wernicke | |
| parameters: | |
| weight: 0.15 # Adds top GPQA performance | |
| density: 0.5 # Sufficient to preserve QA strengths | |
| - model: allknowingroger/QwenStock3-14B | |
| parameters: | |
| weight: 0.15 # For top MMLU-PRO, enhancing domain knowledge | |
| density: 0.5 # Balanced integration of diverse subject expertise | |
| base_model: CultriX/SeQwence-14Bv1 | |
| merge_method: dare_ties | |
| parameters: | |
| normalize: true # Ensures parameter scaling compatibility | |
| int8_mask: true # Memory and computational efficiency | |
| dtype: bfloat16 | |
| adaptive_merge_parameters: | |
| task_weights: | |
| IFEval: 1.2 # Emphasize instruction-following and formatting adherence | |
| BBH: 1.2 # Maintain strong performance in challenging reasoning tasks | |
| MATH_Lvl_5: 1.3 # Ensure domain expertise in competitive math problems | |
| GPQA: 1.3 # Leverage graduate-level knowledge capabilities | |
| MuSR: 1.1 # Enhance multistep reasoning on complex tasks | |
| MMLU_PRO: 1.2 # Ensure robust multitask domain understanding | |
| smoothing_factor: 0.2 # Moderate blending for stable integration | |
| gradient_clipping: 1.0 # Prevent over-contribution from any single model | |