EmbeddingScorer._prepare() passes arg of wrong type to examples_to_samples()

[anthonyduong9]

In

delphi/delphi/scorers/embedding/embedding.py

Lines 76 to 83 in 7166b06

	for i, examples in enumerate(record.test):
	samples.extend(
	examples_to_samples(
	examples, # type: ignore
	distance=i + 1,
	**defaults, # type: ignore
	)
	)

, when record.test is list[ActivatingExample] (the type expected by annotations), examples is of type ActivatingExample, so when

delphi/delphi/scorers/embedding/embedding.py

Lines 111 to 131 in 7166b06

	def examples_to_samples(
	examples: list[Example],
	tokenizer: PreTrainedTokenizer,
	**sample_kwargs,
	) -> list[Sample]:
	samples = []
	for example in examples:
	if tokenizer is not None:
	text = "".join(tokenizer.batch_decode(example.tokens))
	else:
	text = "".join(example.tokens)
	activations = example.activations.tolist()
	samples.append(
	Sample(
	text=text,
	activations=activations,
	data=EmbeddingOutput(text=text, **sample_kwargs),
	)
	)
	return samples

runs, it tries to iterate over an ActivatingExample, leading to a TypeError.

[anthonyduong9]

I just realized that in

delphi/delphi/scorers/surprisal/surprisal.py

Lines 88 to 96 in 7166b06

	for i, examples in enumerate(record.test):
	examples = assert_type(list, examples)
	samples.extend(
	examples_to_samples(
	examples,
	distance=i + 1,
	**defaults,
	)
	)

, record.test is expected to be list[list[ActivatingExample]] rather than list[ActivatingExample]. But I think this might be incorrect, as this is due to #99, which fixed type errors, and in the rest of the codebase, record.test is expected to be list[ActivatingExample] (again, expected by type annotations).

If we want both places to expect record.test to be list[ActivatingExample], I can make the changes in my PR and mark it as ready for review.