Compute the CoSENT loss from embeddings. Args: embeddings: List of embeddings labels: Labels indicating the

Computes the dot-product dot_prod(a[i], b[j]) for all i and j. Args: a (Union[list, np.ndarray, Tensor]): The first tensor.

Do a forward pass on every mini-batch of the input features and yield the embeddings.

Computes the euclidean similarity (i.e., negative distance) between two tensors. Handles sparse tensors without converting to dense when poss

Export a quantized ONNX model from a SentenceTransformer, SparseEncoder, CrossEncoder, or MultiVectorEncoder model. This function applie

Route ``features`` through the resolved sub-module pipeline. Resolves the route from ``task`` and ``modality`` (falling back to values stored

Create a StaticEmbedding instance from a model2vec model. This method loads a pre-trained model2vec model and extracts the embedding

Return a 2D mask where mask[a, n] is True iff a and n have distinct labels. Args: labels: tf.int32 `Tensor` with shape [batch_size

Return a 2D mask where mask[a, p] is True iff a and p are distinct and have same label. Args: labels: tf.int32 `Tensor` with shape

Returns a dictionary of the configuration parameters of the module. These parameters are used to save the module's configuration whe

Enable gradient checkpointing for the model.

Return a ``(modality_config, module_output_name)`` for model types that cannot be handled by the general :meth:`infer_modalities` inference pa

Determine which modalities the processor supports by inspecting its attributes or type.

Returns True if the Huggingface accelerate library is available.

Check if the input is a single example rather than a batch. Redeclared with :class:`~typing_extensions.TypeIs` so type checkers narrow the bra

Whether a single input is a text pair, i.e. cross-encoder ``(query, document)`` input.

Load this module from a model checkpoint. The checkpoint can be either a local directory or a model id on Hugging Face. Args:

Computes the manhattan similarity (i.e., negative distance) between two tensors. Handles sparse tensors without converting to dense when poss

Computes the MeanMaxSim score between two collections of multi-vector embeddings: :func:`maxsim` divided by each query's real token count, wi

Computes the pairwise MeanMaxSim score for each query-document pair: :func:`maxsim_pairwise` divided by each query's real token count, with i

Computes the absolute normalized angle distance. See :class:`~sentence_transformers.sentence_transformer.losses.AnglELoss` or https://hugging

Pool a single ``(num_tokens, D)`` tensor into a ``(num_out, D)`` tensor.

Do forward pass on all the minibatches of the input features and yield corresponding embeddings.

Preprocess the dataset by optionally lazily adding a dataset name column, required for multi-dataset training with multiple losses or

Property to get the processor that is used by this model

Raise a clear ``ValueError`` explaining why ``modality`` is not supported. Shared by :meth:`BaseModel.preprocess` and :meth:`Transformer.preproce

Save a model, then update its config_sentence_transformers.json, e.g. to add requirements.

Performs semantic search using the usearch library. Rescoring will be performed if: 1. `rescore` is True 2. The query embeddings are

.. deprecated:: `tokenize` is deprecated. Use `preprocess` instead.

Whether ``loss`` defers its backward pass to a hook on the loss tensor it returns. Such a loss re-embeds each mini-batch during the *backward* pa

Validate (i.e. check if the dataset IDs exist on the Hub) and process a list of dataset dictionaries. Args: dataset_list

PListMLE loss for learning to rank with position-aware weighting. This loss function implements the ListMLE ranking algorithm which u

CSRLoss implements a combined loss function for Contrastive Sparse Representation (CSR) models. This loss combines two components:

Initialize a BaseModel instance. Args: model_name_or_path (str, optional): If a filepath on disk, loads the model from t

Patch up modules loaded from save formats that predate :class:`MultiVectorEncoder` (PyLate v3, Stanford-NLP ColBERT). Each step is a no-op for

Build a token-level :class:`~sentence_transformers.base.modules.dense.Dense` from a Stanford-NLP ColBERT checkpoint. Stanford-NLP che

Calculate average FLOPS between queries and corpus. FLOPS is the dot product of the average query and corpus count vectors, and can b

Calculate the weight value based on the current step and scheduler type.

Call a :class:`ProcessorMixin` processor, handling both legacy and v5 calling conventions.

Call a single-modality processor (tokenizer, feature extractor, image/video processor).

Whether the active chat template (the processor's, else the tokenizer's) uses a ``task`` variable, e.g. to append query augmentation tokens fo

Run ``model`` over row chunks of ``features`` and return one output dict per chunk. A single-element result means the batch ran in one forward, i

Compute the stats entry for a single column using bounded-memory chunking. Iterates at most 100 rows in chunks of 8 so peak memory is

Runs evaluation during the training

Runs evaluation during the training

Generate an inline HTML tag for a saved example asset.

Zero out prompt token positions in the attention mask so they are excluded from pooling. Operates on a copy: the gradient-cached losses re-em

Human-readable byte formatting.