Check if a value is a single chat message: a dict with ``"role"`` and ``"content"`` keys.

Load projection weights previously written with `save_projection`. If the loss was constructed without a `projection_dim`, the projection lay

Load a Transformer model for the given architecture and task. Args: arch: Architecture name from TINY_MODEL_MAPPING transformer_t

One-shot MaxSim heatmap for a (query, image-document) pair. Args: image: PIL image, URL, or local file path. query_embedding: ``(

Prepend a system prompt to message format inputs. Args: messages: List of message lists (each message list represents one input).

Saves a model and its configuration files to a directory, so that it can be loaded again. Args: path (str): Path on disk

Performs semantic search using sparse embeddings with Qdrant. Args: query_embeddings: PyTorch COO sparse tensor containing query emb

Compute the similarity between two collections of embeddings. The output is a matrix with the similarity scores between all embedding

Check if two sparse embeddings are close to each other. This function works with sparse embeddings in either: 1. Tensor format (assuming

Greedily pack items into ``(start, end)`` chunks whose padded cost of ``(per_item_token * chunk_width + per_item_fixed) * chunk_size`` elements st

Convert legacy ``pooling_mode_*`` bool keys to a single ``pooling_mode`` key in-place. Old keys are removed from *kwargs*. If ``pooling_mode`` is

Converts the input data to a tensor with a batch dimension, stacking lists as :func:`_convert_to_tensor` does. Args: a (Union[li

Format a dataset example value for the model card table, saving non-text values as assets. Delegates the actual file I/O to :meth:`_save_asse

Get the number of samples in sentence features, handling both padded and flattened inputs. With padded inputs, the batch size is the first dimens

Retrieves the model_type from the config_sentence_transformers.json file. This is used to determine whether the model being loaded m

Get the prompt string for a specific column.

Returns the correct learning rate scheduler. Available scheduler: - constantlr, - warmupconstant, - warmuplinear,

Estimate the "size" of an input sample for length-based batch sorting. The exact value doesn't matter, it's only used to group similarly size

Converts the dense tensor to sparse COO if exactly one of the two inputs is sparse, so that mixed inputs can reuse the all-sparse computation

Resolve a prompt from a prompt name or the default prompt name. Args: prompt: An explicit prompt string, or None. pro

Resolve the router mapping for this batch, handling nested (per-dataset) mappings.

Gathers a tensor from each distributed rank into a list, retaining gradients for the local rank's tensor. Args: tensor (torch.Tensor

Decorator for :class:`CrossEncoder.predict` / :class:`CrossEncoder.rank` that handles deprecated keyword arguments. Handles the following legacy

Run the three-step GradCache forward pass. See the module docstring.

Create an SparseStaticEmbedding module from a JSON file containing token to IDF weight mappings. Args: json_path (str):

Returns the name of the device where this module is running on. This function only supports single device or basic distributed training setu

Returns the number of dimensions in the output of :meth:`SparseEncoder.encode`. Unlike :class:`~sentence_transformers.sentence_trans

Import a dotted module path and return the attribute/class designated by the last name in the path. Raise ImportError if the import failed.

Return the row positions for ``ids``, raising ``KeyError(missing_id)`` on a miss.

Normalizes the embeddings matrix, so that each sentence embedding has unit length. Args: embeddings (Tensor): The input embeddings m

Prepend a prompt to text format inputs. For single texts, prepends the prompt directly. For text pairs (cross-encoder inputs), prepen

Computes the cosine similarity between two tensors. Args: a (Union[list, np.ndarray, Tensor]): The first tensor. b (Union[li

Return the slice into ``encode_query``'s output that selects the real content tokens. Chat-template prefixes (e.g. ``<bos>``), ColBERT query mark

Overlay a 2D similarity map onto an image with the mako colormap. Args: image: PIL image, URL, or local file path (loaded via :func:`tran

Performs semantic search using sparse embeddings with Seismic. Args: query_embeddings_decoded: List of query embeddings in format [[

Call ``apply_chat_template`` with the kwarg routing the processor type expects: nested kwargs for a ProcessorMixin, flat with the size kwargs

Convert a relative asset path to a Hub URL if model_id is available, otherwise keep relative.

Return ``(encoder_params, decoder_base_params)`` keyed by name, for the shared base models.

Shared logic for building train/eval/test DataLoaders. Args: dataset: The dataset to build a DataLoader for. batch_si

Overwrite the scores of documents flagged in ``empty`` (no unmasked token), whose every token sat at the masked-fill dtype minimum and whose sum t

Match a mask's token span to the padded embeddings. The chunk loops pass ``truncate=True``, since a batch-wide mask legitimately spans more than o

Generate a usage snippet for non-text inputs (multimodal dicts or single-modality items).

Validate that ``method_output_name`` is present in the method's return type annotation. Returns the name if found, or ``None`` if the method'

Interleave a largest-to-smallest sorted index array so that each consecutive batch contains a mix of long and short inputs. When text

Check if a string is a URL or local file path with one of the given extensions.

Loads a full model using the modules.json file. Args: model_name_or_path (str): The name or path of the pre-trained mode

Build a GISTEmbedLoss without loading real models. ``forward`` encodes via ``self.model`` / ``self.guide``. We replace both with a fake that

Build a CachedGISTEmbedLoss without loading real models. ``calculate_loss`` only needs the configuration attributes and operates on precomput

Create a small dummy PIL image.

A Stanford-NLP ColBERT-shaped checkpoint: ``architectures: ["HF_ColBERT"]``, the projection stored at the repo root as ``linear.weight`` rather th

Derive a ``(B, T)`` mask from a pre-padded ``(B, T, D)`` token-embedding tensor by treating all-zero rows as padding. Real token embeddings n

Whether this pooling applies to inputs encoded for ``task`` (``None`` counts as ``"document"``).

Compute MaxSim scores for knowledge distillation. The query embeddings have shape ``(batch_size, q_tokens, dim)``. The document embeddings have t

Compile the model's forward pass with :func:`torch.compile` to speed up inference. All arguments are forwarded to :func:`torch.compi

Compute the embedding-distillation loss from already-computed embedding lists. Exposed so subclasses (or callers composing this with another

Create test pair samples for all supported modality combinations. Generates pairs for: 1. Text pairs: (text, text) - handled natively by the