Mistral-7B-Instruct-v0.1 identifies Mistral-7B-v0.1 as its parent. Mistral's model card describes the newer model as an instruction fine-tune, trained using public conversation datasets. That named relationship gives an investigator somewhere to look before either model's responses have been tested.
CNIL, France's data-protection authority, uses this example to explain Genmod, its experimental map of open-weight model ancestry. Its 26 August 2026 update adds an English interface and weekly database rebuilds, alongside faster exploration and clearer progress information. The privacy question behind the project is whether personal data memorised by a model could persist when other developers reuse it.
Following a declared relationship
Genmod builds connections from public Hugging Face records. LINC, CNIL's innovation laboratory, explains that publishers supply much of this information through declarations. The graph lets researchers follow ancestors and descendants, but a connecting line does not establish that two models contain the same personal data or reproduce the same response.
LINC illustrates the investigation with a fictional person whose address and birth date appear in a chatbot answer. Looking up the underlying model could identify relatives worth examining. The example is a proposed research scenario, not a reported disclosure involving the Mistral models.
A missing connection has limits too. If a publisher has not recorded a relationship, the graph cannot be treated as an exhaustive account of that model's history. Investigators need evidence about the particular response and model, as well as the lineage declaration.
The rights-request service is still hypothetical
LINC imagines sending a person's query to several suspected models and helping them contact publishers. It labels both features exploratory and hypothetical. One obstacle is that Hugging Face does not require a public contact point for every model publisher. The August update does not announce an operational service for submitting those requests.
For a product team handling a reported disclosure, our recommendation is to record the exact deployed model version, the response under investigation and the declarations used to choose further checks. Keep the graph's displayed refresh date with that record. A future reviewer then has a traceable starting point rather than a screenshot detached from its date.
CNIL's announcement describes improvements to exploration, not a measured guarantee of coverage. We have not tested Genmod's accuracy or run the models. The next task in an investigation remains checking whether a named model reproduces the person's information; a family-tree result cannot answer that by itself.