Verifying LLM Lineage via Model DNA

The public Model DNA approach and Model Genome tool use public artifacts to estimate whether an LLM was trained independently or derived from an existing model, without relying on internal training logs.

The three core signals used in the analysis are:

• Architecture fields in config.json: Compares the simultaneous match of multiple structural values, such as hidden_size, number of layers, number of attention heads, and vocab_size.
• Vocabulary overlap in tokenizer.json: Measures how much the token sets of two models share, normalized by the smaller vocabulary.
• Linear CKA of embeddings: Compares relationships between model embedding spaces using rotation-invariant representation similarity, which can be reproduced with PyTorch.

Rather than asserting lineage based on a single field or metric, the method judges a high likelihood of derivation when multiple signals match simultaneously. However, due to limitations such as the gray area between continued pretraining and full new training, sensitivity to candidate pools and thresholds, and an analysis biased toward embeddings, the results should be interpreted as lineage estimation labels rather than definitive verdicts.