model-merging

Use Model Merging when you need to: Success Stories : Marcoro14-7B-slerp (best on Open LLM Leaderboard 02/2024), many top HuggingFace models use merging Tools : mergekit (Arcee AI), LazyMergekit, Model Soup Linear (Model Soup) SLERP (Spherical Linear Interpolation) Task Arithmetic TIES-Merging DARE (Drop And REscale) Best for : Simple model combinations, equal weighting Best for : Two models, smooth interpolation Layer-specific SLERP: Best for : Combining specialized skills Best for : Many models, resolving conflicts Best for : Reducing redundancy Fix : Only merge models with same architecture Fix : Use more balanced weights (0.3-0.7 range) Fix : Always benchmark before deploying - Combine capabilities from multiple fine-tuned models without retraining - Create specialized models by blending domain-specific expertise (math + coding + chat) - Improve performance beyond single models (often +5-10% on benchmarks) - Reduce training costs - no GPUs needed, merges run on CPU - Experiment rapidly - create new model variants in minutes, not days - Preserve multiple skills - merge without catastrophic forgetting - Simple weighted average of parameters - Fast, works well for similar models - Can merge 2+ models - Interpolates along sphere in weight space - Preserves magnitude of weight vectors - Best for merging 2 models - Smoother than linear - Extract "task vectors" (fine-tuned -...