[flyingdoubleG]

Summary

IntruderScorer(type="internal") crashes while building its batches:

all_quantile_examples = list(quantiled_intruder_sentences.values())
all_quantile_examples.remove(all_active_examples)

list.remove compares each quantile's list with all_active_examples. Lists of the same length are compared element by element, and ActivatingExample uses the default dataclass __eq__, which compares its tokens/activations tensors and raises

RuntimeError: Boolean value of Tensor with more than one value is ambiguous

With the default sampler every quantile holds n_examples_test // n_quantiles examples, so this happens for every batch whose source quantile is not the first one. A record with 10 quantiles × 5 test examples fails on the second batch.

Change

Select the other quantiles by their key (quantile != active_quantile) instead of removing by equality.

Tests

Added test_intruder_internal_takes_intruder_from_another_quantile to tests/test_scorers/test_intruder_example_embedding.py: 3 quantiles × 2 examples. It checks that _prepare_and_batch returns one batch per non-activating example and that each intruder comes from a different quantile than its batch. It raises the error above on main and passes with this change. ruff check and black --check pass.