| merge_method: della_linear | |
| base_model: CultriX/Qwen2.5-14B-Wernickev3 | |
| dtype: bfloat16 | |
| parameters: | |
| epsilon: 0.015 # Fine-grain scaling for precision. | |
| lambda: 1.6 # Strong emphasis on top-performing models. | |
| normalize: true # Stable parameter integration across models. | |
| adaptive_merge_parameters: | |
| task_weights: | |
| tinyArc: 1.75 # Logical reasoning. | |
| tinyHellaswag: 1.65 # Contextual predictions. | |
| tinyMMLU: 1.8 # Domain knowledge. | |
| tinyTruthfulQA: 2.0 # Prioritize truthful reasoning. | |
| tinyTruthfulQA_mc1: 1.85 | |
| tinyWinogrande: 1.9 # Advanced reasoning and predictions. | |
| IFEval: 2.1 # Instruction-following and multitasking. | |
| BBH: 1.9 # Complex reasoning. | |
| MATH: 2.3 # Mathematical reasoning. | |
| GPQA: 2.0 # Factual QA. | |
| MUSR: 2.1 # Multi-step reasoning. | |
| MMLU-PRO: 1.95 # Domain multitask performance. | |
| smoothing_factor: 0.1 # Smooth blending across benchmarks. | |
| gradient_clipping: | |
| CultriX/Qwen2.5-14B-Wernickev3: 0.9 | |
| djuna/Q2.5-Veltha-14B-0.5: 0.92 | |
| qingy2019/Qwen2.5-Math-14B-Instruct: 0.94 | |
| CultriX/SeQwence-14Bv1: 0.88 | |
| allknowingroger/QwenSlerp6-14B: 0.87 | |
| CultriX/Qwenfinity-2.5-14B: 0.85 | |
| models: | |
| - model: CultriX/Qwen2.5-14B-Wernickev3 | |
| parameters: | |
| weight: 0.25 | |
| density: 0.72 | |
| - model: djuna/Q2.5-Veltha-14B-0.5 | |
| parameters: | |
| weight: 0.22 | |
| density: 0.75 | |
| - model: qingy2019/Qwen2.5-Math-14B-Instruct | |
| parameters: | |
| weight: 0.18 | |
| density: 0.74 | |
| - model: CultriX/SeQwence-14Bv1 | |
| parameters: | |
| weight: 0.15 | |
| density: 0.7 | |
| - model: allknowingroger/QwenSlerp6-14B | |
| parameters: | |
| weight: 0.12 | |
| density: 0.68 | |
| - model: CultriX/Qwenfinity-2.5-14B | |
| parameters: | |
| weight: 0.15 | |
| density: 0.7 | |
| tokenizer_source: CultriX/Qwen2.5-14B-Wernickev3 | |