| models: | |
| - model: CultriX/Qwen2.5-14B-Wernicke | |
| parameters: | |
| weight: 0.55 # Backbone model for conversational ability and GPQA | |
| density: 0.80 # Retain most critical parameters for stability and strength | |
| - model: VAGOsolutions/SauerkrautLM-v2-14b-DPO | |
| parameters: | |
| weight: 0.20 # High IFEval and MMLU-PRO performance with minimized weaknesses | |
| density: 0.60 # Focus on impactful parameters for specific benchmarks | |
| - model: rombodawg/Rombos-LLM-V2.6-Qwen-14b | |
| parameters: | |
| weight: 0.25 # Enhanced emphasis on reasoning-heavy tasks like MUSR and MATH | |
| density: 0.70 # Retain reasoning-intensive parameters for improved benchmarks | |
| - model: allknowingroger/Qwenslerp2-14B | |
| parameters: | |
| weight: 0.15 # General stabilizer for consistency across all tasks | |
| density: 0.65 # Focus on balance and avoiding redundancy | |
| base_model: Qwen/Qwen2.5-14B | |
| merge_method: dare_ties | |
| parameters: | |
| normalize: true # Ensure parameter scale consistency | |
| int8_mask: true # Optimize for memory and compute efficiency | |
| dtype: bfloat16 | |
| tokenizer_source: Qwen/Qwen2.5-14B-Instruct | |
| adaptive_merge_parameters: | |
| task_weights: | |
| IFEval: 1.0 # Maintain high IFEval performance | |
| MATH: 1.3 # Prioritize reasoning and calculation-heavy tasks | |
| GPQA: 1.1 # Boost factual recall and reasoning accuracy | |
| MUSR: 1.2 # Enhance logical reasoning and factual understanding | |
| MMLU-PRO: 1.0 # Retain consistent knowledge representation | |
| smoothing_factor: 0.15 # Fine-tune blending for stable transitions between tasks | |
| gradient_clipping: 1.0 # Prevent over-contribution from any single model | |