A transform/map function that adds the dataset name to the batch. Args: batch (dict[str, list[Any]]): The batch of data, where ea

Get the model architecture name for any-to-any task.

Get the model architecture name for text-generation task.

Get the model architecture name for sequence-classification task.

Get the model architecture name for fill-mask task.

Get the model architecture name.

Create a Transformer instance and return it with its supported modalities.

Create a Transformer instance and return it with its supported modalities.

Create a Transformer instance and return it with its supported modalities.

Create a Transformer instance and return it with its supported modalities.

Create a Transformer instance and return it with its supported modalities. Returns: tuple: (arch, model, supported_modalities_list)

The underlying transformer model.

100 samples, 5 labels with 20 samples each.

A lightweight BERT Transformer for reuse across tests.

In the CI environment, we use a temporary directory as `cache_dir` to avoid keeping the downloaded models on disk after the test.

Compute the total loss (base loss + regularizers) from the per-mini-batch embeddings.

Cleanup temporary media files after all tests.

Computes the loss for the BaseModel model. It uses ``self.loss`` to compute the loss, which can be a single loss function or a dicti

The transformers ``PretrainedConfig`` of the underlying model. Several integrations (most notably Deepspeed and ``transformers.Trainer``) rea

The underlying model configuration.

Compute the 2D matrix of cosine distances (1-cosine_similarity) between all embeddings.