What problem does it solve? Combining capabilities from multiple fine-tuned models normally requires expensive retraining, and naive weight averaging causes interference and capability loss. This Skill provides mergekit configurations and algorithms to blend models on CPU without retraining. ## Core Features & Use Cases - Multiple Merge Methods: Configure linear, SLERP, task arithmetic, TIES-Merging, and DARE merges with YAML configs covering weight selection, density tuning, and layer-wise slicing. - Evaluation & Benchmarking: Run Open LLM Leaderboard, MT-Bench, MMLU, HumanEval, and GSM8K evaluations with regression testing and sanity checks before deployment. - Production Patterns: Apply real-world merge recipes (Marcoro14-7B-slerp, MoE creation, gradual merging, quantization) for deployment to HuggingFace Hub. - Use Case: Merge a math-specialized Mistral model with a chat model using SLERP at t=0.5, benchmark the result against both parents, and push the merged model to HuggingFace. ## Quick Start Merge the math model WizardMath-7B and the chat model OpenHermes-2.5-Mistral-7B into one model using SLERP and evaluate it on GSM8K.
