What problem does it solve? Combining capabilities from multiple fine-tuned models normally requires expensive retraining; this Skill merges pre-trained models directly in weight space so you can blend math, coding, and chat expertise into one model without GPUs or training runs. ## Core Features & Use Cases - Multiple Merge Algorithms: Configure linear, SLERP, task arithmetic, TIES-Merging, and DARE merges through mergekit YAML configs, including layer-wise and MoE-style merging. - Unsupervised Coefficient Tuning: Select optimal merge coefficients via generation consistency (AdaMMS method) without labeled evaluation data. - Evaluation & Deployment Guidance: Benchmark merged models on Open LLM Leaderboard, MT-Bench, MMLU, HumanEval, and GSM8K, then quantize and publish to HuggingFace Hub. - Use Case: Merge a math-specialized Mistral model with a chat-tuned variant using SLERP at t=0.5, benchmark the result against both parents, and ship a single model that handles both domains. ## Quick Start Merge the models mistralai/Mistral-7B-v0.1 and teknium/OpenHermes-2.5-Mistral-7B using SLERP with equal weighting and evaluate the result on GSM8K.
