| merge_method: della_linear | |
| base_model: CultriX/Qwen2.5-14B-Wernickev3 | |
| dtype: bfloat16 | |
| parameters: | |
| epsilon: 0.012 # Further reduced to ensure ultra-fine parameter scaling for precision. | |
| lambda: 1.4 # Stronger emphasis on significant model contributions. | |
| normalize: true # Balances the parameter integration for stability. | |
| adaptive_merge_parameters: | |
| task_weights: | |
| tinyArc: 1.6 # Prioritizes logical reasoning improvements. | |
| tinyHellaswag: 1.5 # Strengthened contextual understanding and consistency. | |
| tinyMMLU: 1.65 # Enhanced domain knowledge for multitask benchmarks. | |
| tinyTruthfulQA: 1.9 # Maximized for accurate factual reasoning and QA. | |
| tinyTruthfulQA_mc1: 1.7 # Balanced focus for multiple-choice reasoning. | |
| tinyWinogrande: 1.75 # Advanced reasoning and contextual prediction improvement. | |
| IFEval: 1.9 # Instruction-following tasks boosted by multitask contributors. | |
| BBH: 1.7 # Complex reasoning is supported by logical base models. | |
| MATH: 2.1 # Highest priority, focusing on mathematical excellence. | |
| GPQA: 1.8 # Boosted graduate-level QA capabilities. | |
| MUSR: 1.9 # Nuanced multi-step reasoning strengthened further. | |
| MMLU-PRO: 1.8 # Domain multitask performance maximized. | |
| smoothing_factor: 0.1 # Precisely tuned for smooth task-specific blending. | |
| gradient_clipping: | |
| CultriX/Qwen2.5-14B-Wernickev3: 0.86 # Backbone stability with slightly reduced clipping. | |
| CultriX/Qwenfinity-2.5-14B: 0.83 # Consistent multitask integration. | |
| djuna/Q2.5-Veltha-14B-0.5: 0.91 # Strengthened advanced reasoning contributions. | |
| CultriX/Qwen2.5-14B-Broca: 0.85 # Logical reasoning enhancements stabilized. | |
| qingy2019/Qwen2.5-Math-14B-Instruct: 0.93 # Mathematically focused tasks maximized. | |
| CultriX/SeQwence-14Bv1: 0.88 # Generalist multitask support. | |
| sometimesanotion/Qwen2.5-14B-Vimarckoso: 0.89 # Balanced multi-step reasoning contributions. | |
| allknowingroger/QwenSlerp6-14B: 0.87 # Contextual and logical reasoning integration refined. | |
| models: | |
| - model: CultriX/Qwen2.5-14B-Wernickev3 | |
| parameters: | |
| weight: 0.26 # Core backbone for multitask reasoning. | |
| density: 0.7 # Slight increase to preserve critical reasoning parameters. | |
| - model: CultriX/Qwenfinity-2.5-14B | |
| parameters: | |
| weight: 0.23 # Comprehensive multitask performer. | |
| density: 0.65 | |
| - model: djuna/Q2.5-Veltha-14B-0.5 | |
| parameters: | |
| weight: 0.22 # Advanced reasoning support for GPQA and MUSR. | |
| density: 0.72 | |
| - model: CultriX/Qwen2.5-14B-Broca | |
| parameters: | |
| weight: 0.15 # Logical reasoning and factual QA enhancements. | |
| density: 0.65 | |
| - model: qingy2019/Qwen2.5-Math-14B-Instruct | |
| parameters: | |
| weight: 0.18 # Mathematical reasoning priority. | |
| density: 0.73 | |
| - model: CultriX/SeQwence-14Bv1 | |
| parameters: | |
| weight: 0.14 # Generalist multitask backbone. | |
| density: 0.63 | |
| - model: sometimesanotion/Qwen2.5-14B-Vimarckoso | |
| parameters: | |
| weight: 0.12 # Multi-step reasoning tasks contributor. | |
| density: 0.6 | |
| - model: allknowingroger/QwenSlerp6-14B | |
| parameters: | |
| weight: 0.1 # Contextual reasoning improvements. | |
| density: 0.62 | |
| tokenizer_source: CultriX/Qwen2.5-14B-Wernickev3 | |