A Hugging Face community article introduces Model Genome as a way to compare public model artifacts and document possible lineage. Its useful claim is methodological: make the comparison reproducible instead of turning a similarity score into an accusation. The described workflow checks architecture through shape fields in config.json, tokenizer vocabulary overlap, and weights with linear CKA. A linked Model-Genome-Korea space shows the process as an inspectable example.
Fingerprinting public artifacts
For a team choosing an open model, the workflow starts with files that can be pinned and compared. Save the model IDs, revisions, configuration files and tokenizer data before drawing a conclusion. A configuration match can show that two releases share architectural choices. Tokenizer overlap can show shared vocabulary. Neither one records the full training history.
Three signals, different evidentiary weight
The article rejects row-wise embedding cosine as a valid weights comparison because of rotational invariance. It presents linear CKA as rotation- and isotropic-scale-invariant, so CKA can serve as supporting weight evidence. The same article keeps configuration and tokenizer checks primary because CKA does not cleanly distinguish continued pretraining from a from-scratch run.
Examples include tokenizer overlap around 0.38 in one comparison and 1.000 in another, plus an illustrative sweep covering nine Korean organizations. These are examples supplied by the community article, not an independent peer-reviewed study.
Use it for due diligence, not accusations
A close artifact match may justify a deeper licensing and provenance review. It does not prove that a model was derived from another model, and a weak or missing signal does not prove independence. The author frames the tool as a research aid rather than a legal provenance service. Public files can be incomplete, replaced or missing revisions, so the tested artifacts must remain part of the record.
Pin the next comparison
- Run the configuration and tokenizer checks against two public Hugging Face Hub repositories.
- Save revisions, retrieved artifacts and comparison output.
- Label each axis as primary evidence, supporting evidence, unavailable or inconclusive.
- Add linear CKA only after the first two checks and keep it separate from any legal conclusion.
Watch for independent reruns on pinned revisions, clearer model-card and license disclosures, and tests across different architectures and tokenizer families. Until then, Model Genome is best used as a reproducible check that improves due diligence.
Sources: Hugging Face Community Article — Model DNA and Model-Genome-Korea.