cyberneticlibrary
← library

Merge fine-tuned models without retraining

model-mergingskillsetup L3★9,423
Orchestra-Research/AI-Research-SKILLs ↗
What it does

Combine fine-tuned models without retraining

Best for

Creating specialized models by blending domain expertise (math+coding+chat) in minutes without retraining.

Inputs
  • · model A
  • · model B
  • · merge config (YAML)
  • · method (linear|slerp|ties)
Outputs
  • · merged model
  • · combined capabilities
Requires
  • · mergekit
  • · transformers
  • · torch
Preconditions
  • · models compatible (same arch/tokenizer)
  • · disk space for merged
Failure modes
  • · dtype mismatch
  • · layer count mismatch
  • · incompatible architectures
Trust signals
  • · Marcoro14-7B-slerp top Open LLM Leaderboard
  • · TIES resolves sign conflicts
  • · DARE sparsification reduces redundancy
  • · CPU-only merge