Generate a text-only usage snippet.

Resolve the training max-length override for a column: an int applies to every column, a dict is keyed by task.

Hook for subclasses to inject extra defaults into a module's ``__init__`` at load time. Returned kwargs are forwarded to ``module_class.load(

Returns the SpladePooling module if present, or None. Only searches top-level modules.

Check whether a tokenizers normalizer (or sequence of normalizers) includes Lowercase.

Normalize a value into a list of strings for counting.

Loads the transformers or PEFT configuration Args: model_name_or_path (str): The model name on Hugging Face (e.g. 'sentence-trans

Load a save of a different model type by building this class's default modules on top of it. The source's ``config_sentence_transformers.json

Load encoder-only variants for encoder-decoder architectures. Checks :data:`_ENCODER_ONLY_MODELS` for standard mappings and handles a few spe

Load a HF dataset split with an optional subset/config.

Loads the transformers or PEFT model into the `auto_model` attribute Args: model_name_or_path (str): The model name on Hugging Fa

relative-margin GISTEmbedLoss with no gather/contrast, fed precomputed embeddings (#3819).

relative-margin CachedGISTEmbedLoss fed precomputed embeddings (#3819).

Score one chunk of documents against every query, returned as ``(a_batch, b_batch)`` float32. Pulled out of :func:`maxsim` so the chunk loop stays

Score one chunk of aligned query-document pairs, returned as ``(batch,)`` float32. Pulled out of :func:`maxsim_pairwise` so the chunk loop stays r

Merge per-call ``processing_kwargs`` on top of the instance-level ``self.processing_kwargs``. The merge is shallow per top-level key: for eac

Internal method for multi-process encoding. Either ``pool`` or ``device`` (as a list) must be provided. If ``pool`` is ``None`` and ``device`

Internal method for multi-process encoding. Distributes encoding across multiple processes using the provided pool or list of devices.

anchor0's positive has a negative cosine (-0.50). A non-paired candidate (column 1) is *more* similar (-0.49). anchor1 is paired with that candida

anchor0's positive has a negative cosine (-0.50). A non-paired candidate (column 1) is *more* similar (-0.49). anchor1 is paired with that candida

Validate the dict, fill defaults, and return a fully-populated config (or ``None`` for the no-expansion case).

Pad input features to the next power of 2 for HPU graph compatibility.

Stack per-item ``(tokens_i, dim)`` embeddings into a padded ``(batch, max_tokens, dim)``. ``pad_sequence`` copies item by item, which on an accel

Embeddings whose dot-product similarity is exactly `scores`, one per input pair.

Final construction step, called at the end of ``__init__``: fires every module's :meth:`~sentence_transformers.base.modules.Module.on_model_re

Return a usage tip snippet for the push_to_hub PR description. Subclasses can override this to provide model-type-specific example code.

Resolve the per-sample video metadata entries collected by :func:`_unwrap_video`. If no video in the batch carried metadata, the key is dropped s

Rebuild a per-component loss dict around a single gradient-carrying total. The trainer sums a dict-valued loss for its backward pass, but after g

Record the batch-level ``sampling_rate``, raising if samples disagree. Feature extractors accept a single ``sampling_rate`` for the whole batch,

Reference implementation of the historical dict-based iteration logic.

Tokenize one conversation's skeleton (non-system content replaced by ``filler``), untruncated. The render shares ``_apply_chat_template`` wit

Resolve a string-or-per-dataset mapping to a single string for this batch.

Note why the columns are embedded separately, then return None so the caller falls back. Speed only: the per-column path produces the same loss a

Whether to pack variable-length text inputs into a single flat sequence (FA2 unpadding). Only safe for text-only inputs, since :class:`DataCo

Tie the encoder's parameters to the decoder's: each encoder parameter is replaced by the identically-named decoder parameter, so the two share sto

Split a padded ``(B, T, D)`` tensor into a list of ``(t_i, D)`` per-sample tensors. Pass ``attention_mask`` for the general case. Without a mask,

Unwrap dict-wrapped audio or an ``AudioDecoder`` into a raw array, collecting ``sampling_rate``. Passes through unchanged if ``audio_value`` is a

Unwrap dict-wrapped video or a ``VideoDecoder`` into a raw array, collecting ``video_metadata``. Passes through unchanged if ``video_value`` is a

Update the default model ID in the model card.

Validate prompt configuration and log prompt information.

Tokenizer-aware checks. The structural ``_normalize_query_expansion`` runs without the tokenizer, so this method catches the two silent-wrong-

Verify that causal inputs are left-padded, i.e. that every sample ends in a real token. Checks the produced ``attention_mask`` rather than ``

Raise if the chat template cannot carry the ``query``/``document`` roles a pair is mapped to. A template that only branches on ``system``/``u

Whether dataloader workers start via ``spawn`` rather than ``fork``/``forkserver``. Only ``spawn`` pays the import cost per worker. A ``forks

View a numpy array as a tensor without copying its buffer, which on a corpus of embeddings costs as much as the scoring it feeds. Two kinds of arr

Add a callback responsible for automatically tracking data required for the automatic model card generation This method is called in

Creates a new 'notice' logging level

Compute differentiable approximate ranks using the ApproxRank formulation. For each document i: ``approx_rank(i) = 1 + sum_{j != i} sigmoid(a

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

bf16/fp16 autocast for evaluator calls outside the trainer.

bf16/fp16 autocast for evaluator calls outside the trainer (which has its own autocast).

Build the triplet loss over a batch of embeddings. We generate all the valid triplets and average the loss over the positive ones. Arg