Pedigree
An open-source AI model license auditing tool built on a graph database. It traces model lineages, compares license obligations along the lineage, and flags potential inconsistencies.
Timeline (last 90 days)
Pedigree's workflow is: after entering a model ID, it reads the public metadata on the Hugging Face Hub (license, base model, declared training datasets), writes it into the FalkorDB graph database, then traverses upward along the lineage, comparing each child model's declared license against the obligations required by the parent model's license.
Pedigree's design principle is that any missing or unreadable information counts as risk and is never counted as passing.
The base_model field in Hugging Face Hub metadata is filled in manually by the publisher and is not verified by the Hub; the field can point to any model ID or be left blank, so Pedigree audits the declared lineage rather than the actual training source.
Each of Pedigree's flags is essentially a Cypher query: traversing upward along the lineage up to six levels, looking for obligations carried by the parent license but missing from the child license.
Pedigree has so far compared the license texts of Llama 3 and Llama 3.2 clause by clause; rule flags for other licenses are marked as unverified, and flags produced by unverified rules are only leads that may miss or falsely report issues and cannot be used as a basis for legal judgment.
In terms of privacy design, Pedigree only reads public metadata, requires no personal information and has no account system; the browser stores a random ID, review records are kept in a private graph under that ID, a Forget button can delete the entire graph, and no tokens are stored in the repository.
Each time a user saves a decision, the Pedigree system generates a separate private graph and stores a snapshot of the lineage at that time; when a model is re-checked, the old parent links are deleted first.
Pedigree's code is open-sourced on GitHub (triixdev/Pedigree).
All Facts (8)
Pedigree's workflow is: after entering a model ID, it reads the public metadata on the Hugging Face Hub (license, base model, declared training datasets), writes it into the FalkorDB graph database, then traverses upward along the lineage, comparing each child model's declared license against the obligations required by the parent model's license.
50%UnverifiedPedigree's design principle is that any missing or unreadable information counts as risk and is never counted as passing.
50%UnverifiedThe base_model field in Hugging Face Hub metadata is filled in manually by the publisher and is not verified by the Hub; the field can point to any model ID or be left blank, so Pedigree audits the declared lineage rather than the actual training source.
50%UnverifiedEach of Pedigree's flags is essentially a Cypher query: traversing upward along the lineage up to six levels, looking for obligations carried by the parent license but missing from the child license.
50%UnverifiedPedigree has so far compared the license texts of Llama 3 and Llama 3.2 clause by clause; rule flags for other licenses are marked as unverified, and flags produced by unverified rules are only leads that may miss or falsely report issues and cannot be used as a basis for legal judgment.
50%UnverifiedIn terms of privacy design, Pedigree only reads public metadata, requires no personal information and has no account system; the browser stores a random ID, review records are kept in a private graph under that ID, a Forget button can delete the entire graph, and no tokens are stored in the repository.
50%UnverifiedEach time a user saves a decision, the Pedigree system generates a separate private graph and stores a snapshot of the lineage at that time; when a model is re-checked, the old parent links are deleted first.
50%UnverifiedPedigree's code is open-sourced on GitHub (triixdev/Pedigree).
50%