fix(evaluator): preserve multi-alias targets for inclusion-based scoring

[Dhru1001]

Bug

TaskState.target always returns Target.text, which joins every accepted answer alternative into a single string with no separator:

@property
def text(self) -> str:
    return ''.join(self.target)

For a benchmark like trivia_qa, whose samples set target to a list of answer aliases (record['answer']['aliases'] + record['answer']['normalized_aliases']), this collapses e.g. ["Oprah Winfrey", "Oprah", "Winfrey, Oprah Gail", ...] into one unmatchable blob: "Oprah WinfreyOprahWinfrey, Oprah Gail...".

Accuracy(allow_inclusion=True) (used by trivia_qa and mmlu_redux) is explicitly written to accept either a single string or a list of alternatives:

alternatives = [reference] if isinstance(reference, str) else reference

but every call site passes task_state.target, which is always a str - so alternatives is always [the_joined_blob], and a correct prediction can never exact-match it. Every sample in trivia_qa scores 0 regardless of whether the model's answer was right, because there's nothing left to match against.

Confirmed on a real run: the model answered "Of which African country is Niamey the capital?" with ANSWER: Niger (correct), and the sample still scored 0 - the gold reference used for grading was the fully-concatenated alias blob shown above.

Fix

Return the full list of alternatives from TaskState.target when a sample has more than one, instead of always flattening to Target.text. Single-target samples (the common case across the rest of the benchmark suite) are unaffected and still return a plain string, since len(self._target) > 1 is false for them - so this is scoped to exactly the multi-alias case that was broken.

Checked every other call site that reads task_state.target for a str-only assumption (e.g. json.loads(task_state.target) in researchrubrics) - none of them ever set a multi-item target, so they're untouched by this change.

[Dhru1001]

@Yunnglin

[Yunnglin]

Thanks for identifying the multi-alias scoring bug — the underlying issue is real, and preserving aliases for Accuracy(allow_inclusion=True) fixes the scoring path.

However, this currently breaks the predictions API for affected runs:

TaskState.target now returns list[str] → ReviewResult.from_score_state() persists that list → _build_prediction_row() assigns it to Gold, while PredictionRow.gold and the generated frontend contract still require Gold: string.

get_predictions() validates its successful response through PredictionsResponse, so a row with Gold=['A', 'B'] raises a Pydantic validation error and the endpoint returns 500. I reproduced this directly with PredictionsResponse.model_validate().

Could you please keep the serialized/display target as a string and expose aliases through a scoring-specific path, or update the complete API/frontend contract chain to support string | string[]? The former seems more scoped for this fix.

Please also add a regression test covering the full TaskState -> scoring -> ReviewResult -> predictions endpoint path for a multi-alias TriviaQA/MMLU-Redux sample.

[Yunnglin]

LGTM. The multi-reference scoring, cache migration, predictions API, and Dashboard rendering paths are covered by the updated tests and all required checks are green.

[Dhru1001]

Thxxx @Yunnglin