Create two large sparse tensors of shape (50, 100) each.

Returns the chunk size of the SpladePooling module, if present. This chunk size is along the sequence length dimension (i.e., number

Qdrant is not a test dependency, so stand in for the imports that the function performs.

Seismic is not a test dependency, so stand in for the import that the function performs.

2D labels with multi-column input = same target broadcast (multilingual-distillation case).

3D labels supply a different teacher embedding per text column.

4D model outputs should be flattened to 3D.

The results CSV data row must have as many columns as the header row. The ``*_accuracy_threshold`` and ``*_f1_threshold`` metrics contain a secon

The euclidean/manhattan metrics must treat smaller distances as more similar. ``pairwise_euclidean_sim``/``pairwise_manhattan_sim`` return negati

Tests that the Acc score for the computed threshold is correct

Tests that the F1 score for the computed threshold is correct

Tests that the max_* aggregation does not assume that cosine was requested

Tests that the LabelAccuracyEvaluator can be loaded correctly

Tests that the ParaphraseMiningEvaluator can be loaded

Tests that the TripletEvaluator can be loaded & used

AdaptiveLayerLoss deposits ``all_layer_embeddings`` in the per-column feature dicts on the final-layer pass and reads it back on every prior-layer

For nested wrappers the error should name both the outer wrapper (what the user passed in) and the inner object the unwrap got stuck on.

Regression test for #3757: the final-layer embeddings are the (teacher) target of the self-distillation KL loss and must be detached, so no gradie

Each prior layer loss is scaled by the scheme's weight for its layer index, averaged over the number of sampled layers. KL is disabled so the tota

Each prior-layer pass re-runs the pooling half over the same feature dicts. If pooling wrote its prompt-excluded mask back into them, every pass a

If `self.model` is something we can't unwrap to a BaseModel, fail loudly instead of silently raising the cryptic `'X' object is not subscriptable`

Sanity check for the non-DDP, non-compile path: a bare SentenceTransformer skips the unwrap loop entirely.

Compile-of-DDP and DDP-of-compile both need to unwrap to the inner BaseModel.

AdaptiveLayerLoss.forward reaches into `self.model[0]`. Under DDP / torch.compile the trainer rebinds `loss.model` to the wrapper (and binds BaseM

AdaptiveLayerLoss calls its base loss once per layer, which gradient caching cannot support. The ``uses_gradient_cache`` capability check must cat

Renames (sentences→inputs, activation_fct→activation_fn) are wired via deprecated_kwargs. num_workers is removed as a no-op.

When all strategies fail (no source path), RuntimeError is raised.

When forward does not accept ``**kwargs`` (``model_forward_params`` is a set), forward() drops any feature key outside the allowlist.

Anchors (queries) are typically much narrower than the candidate columns, and merging them in would pad every anchor row to candidate width. Expec

Only writes when there's a header plus at least one data row, and then appends every value to that row. Otherwise it no-ops and leaves the file un

Audio array + text pairs should be converted to message format.

Training example table asset paths are rewritten to absolute Hub URLs.

forward should fall back to getattr when dict indexing fails on model output.

AudioDict saved to assets/ as .wav file.