spec.md

Definitions

• <model_name>: The shortcut or identifier name of a pre-trained model available from Huggingface (e.g. bert-base-uncased, roberta-large).
• <config_id>: 16-char hash of the adapter architecture.
• <config>: an identifier for an adapter architecture (e.g. pfeiffer).
• <task>: the name of the task the adapter is categorized under (e.g. sentiment, question_answering, nli for task adapters or en, de for language adapters).
• <subtask>: the dataset or domain the adapter was trained on (e.g. multinli, squad1.1, wiki)
• <weights_name> (of an adapter): the name under which adapter weights are saved in the weights file. Given in the name attribute of the adapter config and required for reloading. (This name originates from calling model.add_adapter(<weights_name>)).
• <type>: the adapter type (e.g. text_task).
• <org_name>: name of the orga., GH user... that maintains the adapter.

Loading

Adapters are loaded with: model.load_adapter(<specifier>, config=<config>, model=<model_name>, version=<version>).

<specifier> can be one of the following iff globally unique:

• <task>
• <task>/<subtask>
• <subtask>
• <task>@<org_name>
• <task>/<subtask>@<org_name>
• <subtask>@<org_name>

File structure

Index files

Placed in /dist/index/<model_name>.json.

{
    <task>: {
        <subtask>: {
            "default": "<org_name>/<file_name>.json",
            <config_id>: {
                "default": "<org_name>/<file_name>.json",
                "versions": {
                    <org_name>: "<org_name>/<file_name>.json"
                    ...
                }
            }
            ...
        }
        ...
    }
    ...
}

Config files

Placed in /adapters/<org_name>/<file_name>.yaml.

Schema defined in /scripts/schemas/adapter.schema.json.