Model Genome: Fingerprinting Whether an LLM Was Trained from Scratch or Derived
Model Genome offers a method to determine if LLMs were trained from scratch or derived.
When a lab claims to have developed a model from scratch, how can this be verified? We created a reproducible pipeline to fingerprint models across three axes: architecture, tokenizer, and weights. This method was applied to the public foundation models of nine Korean organizations, revealing a spectrum of originality in model development.