| merge_method: dare_ties # Changed to dare_ties | |
| base_model: CultriX/Qwen2.5-14B-Wernickev3 | |
| dtype: bfloat16 # Use float32 for maximum precision. | |
| out_dtype: bfloat16 # Output model also uses bfloat16 for consistency and reduced memory usage. | |
| parameters: | |
| t: 0.5 # Balances interpolation between models; 0.5 gives equal weight to all contributors. | |
| normalize: true # Ensures parameters are normalized to maintain stability during merging. | |
| rescale: true # Aligns parameter scales across models for better integration. | |
| int8_mask: false # Disable int8 masking to preserve full precision during merging. | |
| epsilon: 0.008 # Ultra-fine parameter scaling for precise adjustments between models. | |
| lambda: 1.8 # Emphasizes high-impact parameters, giving more weight to significant contributors. | |
| adaptive_merge_parameters: | |
| task_weights: # Assign weights to tasks based on their priority and impact on benchmarks. | |
| tinyArc: 1.6 # Logical reasoning benchmark; slightly lower priority. | |
| tinyHellaswag: 1.5 # Contextual reasoning benchmark with moderate priority. | |
| tinyMMLU: 1.8 # Multi-domain knowledge benchmark; important for multitask performance. | |
| tinyTruthfulQA: 1.9 # Focuses on factual reasoning and QA; high priority. | |
| tinyTruthfulQA_mc1: 1.75 # Multiple-choice factual reasoning; closely related to TruthfulQA. | |
| tinyWinogrande: 1.75 # Core reasoning benchmark; slightly lower than BBH. | |
| IFEval: 2.30 # Instruction-following tasks; given a high priority for practical applications. | |
| BBH: 2.05 # Complex reasoning benchmark; critical for logical tasks. | |
| MATH: 2.70 # Highest priority to emphasize mathematical reasoning excellence. | |
| GPQA: 2.20 # Graduate-level QA tasks; balanced priority for high-level reasoning. | |
| MUSR: 2.15 # Multi-step reasoning; slightly increased to strengthen reasoning performance. | |
| MMLU-PRO: 2.00 # Domain multitask benchmark; maintained for general multitask capability. | |
| smoothing_factor: 0.03 # Low smoothing for precise task-specific blending without over-generalizing. | |
| gradient_clipping: # Control gradient clipping for each model to stabilize training. | |
| CultriX/Qwen2.5-14B-Wernickev3: 0.89 # Higher value ensures stability for the base model. | |
| djuna/Q2.5-Veltha-14B-0.5: 0.92 # Stable setting to enhance reasoning contributions. | |
| CultriX/SeQwence-14B-EvolMerge: 0.87 # Moderate value for generalist multitask support. | |
| qingy2024/Fusion4-14B-Instruct: 0.93 # High stability to emphasize mathematical tasks. | |
| CultriX/Qwen2.5-14B-Emerged: 0.88 # Stable setting to maintain multitask performance. | |
| sometimesanotion/Lamarck-14B-v0.6: 0.89 # Stable contribution for multi-step reasoning. | |
| allknowingroger/QwenSlerp6-14B: 0.90 # Adjusted for stable integration of the replacement model. | |
| hotmailuser/QwenSlerp2-14B: 0.91 # Increased slightly for stable integration of reasoning contributions. | |
| models: # Define models to include in the merge, along with their weights and densities. | |
| - model: CultriX/Qwen2.5-14B-Wernickev3 | |
| parameters: | |
| weight: 0.33 # Increased to absorb some of the weight from the removed model. | |
| density: 0.78 # Maintained optimal density for robust generalist performance. | |
| - model: djuna/Q2.5-Veltha-14B-0.5 | |
| parameters: | |
| weight: 0.28 # Increased slightly to enhance reasoning benchmarks like MUSR. | |
| density: 0.77 # Maintained for strong nuanced reasoning. | |
| - model: allknowingroger/QwenSlerp6-14B # Replacement for Qwenfinity-2.5-14B. | |
| parameters: | |
| weight: 0.15 # Matches the weight of the replaced model to preserve balance. | |
| density: 0.70 # Increased slightly for stronger parameter integration. | |
| - model: CultriX/SeQwence-14B-EvolMerge | |
| parameters: | |
| weight: 0.12 # Moderate weight for general multitask support. | |
| density: 0.62 # Maintained for stable contribution. | |
| - model: qingy2024/Fusion4-14B-Instruct | |
| parameters: | |
| weight: 0.09 # Moderate weight; focuses on mathematical reasoning tasks. | |
| density: 0.75 # Maintained density for stable integration. | |
| - model: CultriX/Qwen2.5-14B-Emerged | |
| parameters: | |
| weight: 0.08 # Balanced weight for multitask contributions. | |
| density: 0.69 # Maintained density for stable integration. | |
| - model: sometimesanotion/Lamarck-14B-v0.6 | |
| parameters: | |
| weight: 0.06 # Lower weight to allow more impactful models to dominate. | |
| density: 0.62 # Maintained for stable multi-step reasoning contribution. | |
| - model: hotmailuser/QwenSlerp2-14B | |
| parameters: | |
| weight: 0.11 # Increased slightly to balance contributions. | |
| density: 0.66 # Maintained for stable parameter integration. | |