Reset model_card_data fields to ensure a clean snippet test.

Set up common mocks for push_to_hub tests. Returns a dict with: - "upload_folder": dict capturing the most recent upload_folder kwargs

Evaluate the model based on an evaluator Args: evaluator (BaseEvaluator): The evaluator used to evaluate the model.

Save the PyTorch weights of the module to disk. Args: output_path (str): The path to the directory where the weights sho

Re-raise ImportError/AttributeError with an install hint when a multimodal dependency is missing.

Convert a typed input dictionary to message format. Produces a single message with the given ``role``. For query/document pairs, use

Derive one conversation's trailing chat-template suffix (cached). The conversation is rendered twice with different fillers: the longest comm

Compute an identity hash for an asset value, or None if the type is not supported.

Create a ``_MockVideoDecoder`` with custom metadata overrides.

Convert a value to a format suitable for model inference. ``VideoDecoder`` objects are converted to a ``VideoDict`` via :meth:`_video_decoder

Process chat messages using the processor's chat template. ``chat_template_kwargs`` is forwarded to ``apply_chat_template``, except the Sente

Restore the trailing chat-template suffix that truncation may have removed (in place). Chat templates append fixed tokens after the content (

The dimensionality of each token vector returned by :meth:`encode`.

Check if a string is an audio URL or file path.

Load the configuration of the module from a model checkpoint. The checkpoint can be either a local directory or a model id on Hugging Face.

Normalize messages to the target format (``self.message_format``). Extra keys beyond ``"role"`` and ``"content"`` are preserved during conver

Convert ``skiplist_words`` to token IDs using ``tokenizer``. Called by :class:`MultiVectorEncoder` after the tokenizer is fully initialised.

Sets the ``include_prompt`` attribute in the pooling layer in the model, if there is one. This is useful for models where the prompt

Compute the all-pairs score matrix between two collections of multi-vector embeddings, using this model's :attr:`similarity_fn_name`.

Count images and videos per sample from the message structure. Some VLM processors flatten per-sample visual tokens into single tensors (e.g.

Format a value for inclusion in a code snippet. Strings are shown as repr (quoted), and asset paths are converted to Hub URLs when mo

Infer the message format expected by the processor. Checks known model types first, then inspects the processor's chat template for p

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

Convert a ``VideoDecoder`` to a ``VideoDict`` by extracting all frames. Tries multiple strategies to extract frames: 1. ``get_frames_

Adds a fresh new adapter to the current model for training purposes. If no adapter name is passed, a default name is assigned to the

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

Format a modality for display, e.g. ``("text", "image")`` becomes ``"text+image"``.

Chat-template mock: one token per content character plus ``suffix``, honoring truncation/max_length and return_tensors ("pt", "np", or ragged list

Modality kwargs for _process_chat_messages with only the text bucket populated.

Per-query-token similarity over a 2D image-patch grid. Args: query_embedding: ``(Qt, D)`` per-token query embeddings. image_embed

Computes the pairwise cosine similarity cos_sim(a[i], b[i]). Args: a (Union[list, np.ndarray, Tensor]): The first tensor. b

Performs ranking with the CrossEncoder on the given query and documents. Returns a sorted list with the document indices and scores.

Determine whether text-only inputs can be flattened (concatenated without padding) for more efficient inference. When enabled, text-only inpu

Converts like :func:`_convert_to_tensor`, then upcasts sub-float32 floats (fp8, float16, bfloat16) to float32: matmul rounds its output to th

Check if a dict is an AudioDict or VideoDict (vs a multimodal input dict).

Prepend a prompt string to each text input.

The rank of the current process in the distributed group, or ``0`` when not distributed.

Compute the intersection of two sparse embeddings via element-wise multiplication. For each dimension, the result retains the minimu

Check if the code is running in a Continuous Integration (CI) environment. This is determined by checking for the presence of certain environ

Returns whether ``torch.distributed`` is available and has been initialized. The availability check must come first: some PyTorch builds (e.

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

Computes the euclidean distance (i.e., negative distance) between pairs of tensors. Args: a (Union[list, np.ndarray, Tensor]): The f

Computes the manhattan similarity (i.e., negative distance) between pairs of tensors. Args: a (Union[list, np.ndarray, Tensor]): The

Resolve the route from ``task`` and ``modality``, then delegate preprocessing to the first module of the matched route. The returned dictionar

Persist the projection layer's weights so they can be reused in a later run. The standard Trainer save path only writes the student model, no

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

Check if the model supports the given modality. A modality is supported if: 1. It is directly listed in :attr:`modalities` (includin

Converts a similarity function name or enum value to the corresponding similarity function. Args: similarity_function (U

Converts a similarity function into a pairwise similarity function. The pairwise similarity function returns the diagonal vector fro

Convert padded inputs to flattened format using DataCollatorWithFlattening.

Check if the first module (Transformer or Router containing Transformers) supports flattened text-only inputs. For Router models, this return

Compute similarity between a single query and passage using the model. This helper mirrors the internal scoring logic in `mine_hard_negatives` by

Extract the output field names from a method's return type annotation. Args: method (Callable): The method to inspect. R

Return the length of the prompt in tokens, excluding any trailing special token. Returns None if the processor does not produce ``input_ids``

Build the kwargs dict for ``__init__`` by merging config file, hub kwargs, and caller overrides. Priority (highest to lowest): caller kwargs

Load a module class from a class reference string, preferring a class supplied via ``module_classes``. Args: class_ref:

Parse model configuration and merge saved prompts/defaults with user-provided values. User-provided prompts and default_prompt_name take prec

Resolve the route key based on task and modality. Args: task: The task type (e.g., "query", "document") moda

ColBERT (MaxSim) contrastive scoring for in-batch negatives. Takes ``(Q_query, q_tokens, dim)`` query embeddings and ``(Q_doc, N, d_tokens, dim)`

Create test samples for all supported modalities and their combinations. Generates samples for: 1. Each single modality with different input

Deprecated training method from before Sentence Transformers v4.0, it is recommended to use :class:`~sentence_transformers.cross_enco

Returns the version of an installed package, or None if the package can't be found. The ``__version__`` of an already imported module takes

Get the keyword arguments specific to this model for inference methods like `encode` or `predict`. Example: >>> from se