Model merging and model families

In short: Several fine-tuned models of the same origin can be fused into one without retraining. Models also appear in families of several sizes and variants.

In more detail: Because fine-tuned models from the same starting point have only small weight changes, these changes can be combined.

In detail

• Weight averaging: the average of several models (“model soups”) is often better than each single one.
• Task arithmetic: compute the “task vector” (fine-tuned minus base) and add or subtract it.
• TIES, DARE, SLERP: methods that reduce conflicts between changes.
• Tool: mergekit, popular in the open-model community.
• Families: a provider often releases sizes (1B, 8B, 70B), variants (base, instruct, code, vision) and quantisations (Quantisation).
• Naming: Name-8B-Instruct-Q4_K_M = 8 billion parameters, chat variant, 4-bit quantisation.
• Licence: merged models inherit all licences of the ingredients, see Open models and their licences.

See also: Fine-tuning and LoRA, Hugging Face