On the Hugging Face Hub, the base_model field in a model card names the model or models a repository was derived from. The Hub infers the relationship as finetune, adapter, merge or quantized. For supply chain review, that lineage tells you which upstream licenses and risks you are inheriting.
What the field records
The Hub docs say that if a model is a fine-tune, an adapter, or a quantized version of a base model, you can specify the base model in the card metadata. It can be one model ID or a list. The Hub uses it to show the base model on the page and to let users filter for derived models. A base_model_relation field can set the relationship explicitly, with quantized given as an example value.
base_model:
- org-a/model-one
- org-b/model-two
base_model_relation: merge
The relationship types the Hub infers are adapter, merge, quantized and finetune.
What each type means for you
Finetune. Weights trained further from a parent. The parent's license terms may still apply, and any problem in the parent can carry through.
Adapter. Small add-on weights that attach to a base model, such as LoRA or PEFT. You need both pieces. Review the adapter and the base separately, and pin both.
Merge. Weights blended from two or more parents. Each parent can bring its own license terms, so read all of them and get legal advice where it matters.
Quantized. A smaller copy of the same model. Quantizing is a conversion step done by someone who is not necessarily the original author, so file format and integrity of that copy deserve their own check.
Limits of the field
The metadata is optional and author supplied. A missing base_model does not prove the model is original, and a present one can be wrong. Where lineage matters, such as for a license decision, ask the author or compare with the upstream project's own statement.
Using lineage in your review
- Record every parent in your inventory entry.
- Fetch each parent's license text.
- Check each parent for known issues and update status.
- Pin the exact revision of every piece you load.
- If the lineage is unclear, treat the model as unreviewed.
For adapters and merges in more depth, see LoRA adapters and model merges: who do you trust. For the card as a whole, see model card fields to read.
An example walk
Imagine a quantized model whose card says it is based on a fine-tune, which in turn lists a base model. Three licenses, three upstream histories, and a conversion step done by a third party. A reviewer who stops at the first card sees one project. One who follows the base_model links sees the real chain, and can ask the right question at each hop. Following the chain takes minutes and it is where license surprises usually appear.
Quick check
Does base_model prove where a model really came from?
No. It is author supplied metadata. Use it as a lead, then confirm with upstream sources.
Why record every parent of a merge?
Each parent can carry its own license terms and issues, and your model inherits all of them.
Sources
Last reviewed: 19 October 2026.
