`BoolQ` ground-truth label appears inverted

[HuggingNobody]

In evaluation/benchmarks.py:322-330:

class BoolQ(HFMCQBenchmark):
    def __init__(self, split: str = "validation", num_shots: int = 5):
        super().__init__(
            path="google/boolq", split=split, shot_split="train", num_shots=num_shots,
            row_to_doc_fn=lambda r: MCQDoc(
                f"{r['passage']}\nQuestion: {r['question'].capitalize()}?",
                ["Yes", "No"], 1 - int(bool(r['answer']))
            )
        )

The google/boolq dataset has answer: bool where True means the correct answer to the question is "yes". With choices ["Yes", "No"], the gold index for True should be 0, not 1 - 1 = 0… wait, 1 - int(True) == 0, so True → 0 → "Yes". That's correct.

But for False: 1 - int(False) == 1 → "No". Also correct.

OK — false alarm on the label inversion, sorry. However, there's a real bug nearby: r['question'].capitalize() will lowercase every character after the first one, so a question containing acronyms like "Was DNA discovered..." becomes "Was dna discovered...". Capitalize is destructive here; a simple r['question'][0].upper() + r['question'][1:] would preserve case correctly.

Severity: Low (changes the prompt distribution, may slightly hurt benchmark scores).