model-merging

Write a YAML config specifying a merge_method such as slerp or linear, list the source models with weights or an interpolation factor t, then run mergekit-yaml config.yml ./merged-model --cuda. The merged model loads with standard transformers APIs.

Use SLERP for smoothly blending two models, linear averaging for a few similar models, and Task Arithmetic or TIES-Merging when combining multiple task-specific models. DARE or DARE-TIES reduces redundancy and often performs best when merging three or more models.

No, merging only works between models sharing the same architecture, such as two Mistral-7B variants. Mismatched architectures like Llama and Mistral have incompatible weight layouts and cannot be merged.

Use generation consistency: merge with several candidate coefficients, generate responses on a small unlabeled prompt set, and pick the coefficient whose outputs are most similar to its neighbors. This unsupervised proxy from AdaMMS (arXiv:2503.23733) avoids manual grid search.

Common causes include over-weighting one model beyond the 0.3-0.7 range, merging incompatible architectures, or excessive sparsity in TIES/DARE density settings. Always benchmark the merged model on MMLU, GSM8K, or MT-Bench before deploying.