An architecture is a model's skeleton and a checkpoint is its trained weights. Adapters, merges and quantized versions are all derived from a base model. Cards document models and datasets, and an ML BOM is the inventory that ties them together. Knowing the terms keeps your records accurate.
Core terms
Architecture. The skeleton of a model, the layers and operations. Hugging Face's loading guide uses BERT as an example of an architecture.
Checkpoint. The weights for a given architecture. The same guide gives google-bert/bert-base-uncased as a checkpoint of BERT. When people say "download the model" they usually mean a checkpoint.
Weights file. The file that stores a checkpoint, in a format such as pickle based PyTorch files or safetensors. The format decides what risk loading it carries.
Sharded checkpoint. A large checkpoint split into several files, with an index file that maps each parameter to the shard that holds it.
Derived models
The Hub's model card documentation lets a card name a base model and describes the relationship types:
| Term | Meaning |
|---|---|
| Finetune | A base model trained further on new data |
| Adapter | A small add on, such as LoRA or PEFT, used with a base model |
| Merge | A model built from two or more existing models |
| Quantized | A smaller, lower precision version of a base model |
For supply chain records this matters because the risk and the license travel with the parent. A quantized copy published by a stranger is two trust decisions: the original and the person who converted it.
Documentation terms
Model card. A README with metadata that describes a model, its intended uses and limits, training information, datasets and results.
Dataset card. The same idea for data, covering contents, license, language, size and known biases.
Metadata block. The YAML section at the top of a card. Tools use it to filter and display licenses, base models and datasets.
Inventory terms
SBOM. A software bill of materials, the list of components in software.
ML BOM. CycloneDX describes its ML BOM capability as representing information about models, datasets and their dependencies, including dataset provenance and configurations. SPDX 3.0 handles similar ground through its AI profile and the AIPackage class.
Provenance. Where something came from and how it got to you. For a model: who published it, which revision, and what it was built from.
Revision. A specific version in a repository, ideally a full commit hash, which does not move the way a branch does.
Why these terms are worth agreeing on
Teams often write "model" in a record and mean three different things. One person means the architecture, another the checkpoint they downloaded, a third the adapter they trained on top. Agree on a short vocabulary, then put each part in its own row of the inventory with its own source and license.
A simple rule: one row per thing you can download, one link from each derived item to its parent.
Quick check
Is a LoRA adapter a separate supply chain component?
Yes. It has its own publisher and license, and it depends on a base model. Record both and the link between them.
Does a quantized model inherit the base model license?
The base model's terms still matter, and the converter may add terms of their own. Read both before you use it.
Safeguard scans dependencies and generates SBOMs for your repositories. Try a free scan at gold.safeguard.sh/scan.
Sources
- Transformers loading models guide
- Hugging Face Hub model cards
- Hugging Face Hub dataset cards
- CycloneDX ML BOM capability
- SPDX 3.0.1 AIPackage class
Last reviewed: 19 October 2026.
