PassAtK returns 1.0 for all-wrong samples when k > n #1410

PassAtK.pass_at_k returns 1.0 whenever n - c < k (metrics_sample.py:1321). With k > n that holds for every input, so a sample with zero correct answers scores a perfect pass@k. GPassAtK in the same file returns 0.0 for k > n (_compute_g_pass_at_k, metrics_sample.py:1416).

Repro on main (b634749), Python 3.12.13, fresh venv, pip install -e .:

from lighteval.metrics.metrics import Metrics
from lighteval.metrics.metrics_sample import GPassAtK
from lighteval.tasks.requests import Doc
from lighteval.models.model_output import ModelResponse

doc = Doc(query="2+2?", choices=["4"], gold_index=0)
resp = ModelResponse(text=["5", "6", "7", "8"])  # 4 samples, all wrong

for k in (2, 4, 5, 10):
    m = Metrics.pass_at_k(sample_params={"k": k, "n": 4})
    print(f"pass@k k={k} n=4:", m.sample_level_fn.compute(doc, resp))

g = GPassAtK(k=10, n=4, strip_strings=True)
print("g-pass@10 n=4:", g.compute(doc, resp)["g-pass@10_0.0"])

pass@k k=2 n=4: 0.0
pass@k k=4 n=4: 0.0
pass@k k=5 n=4: 1.0
pass@k k=10 n=4: 1.0
g-pass@10 n=4: 0.0

k and n come from sample_params (the same call the built-in tasks use, e.g. math_500.py:69), and nothing checks k <= n. Since only the corpus mean is reported, a config like {"k": 16, "n": 8} shows up as a pass@16 of 1.0 with no warning.

Ways to fix it:

1. Raise ValueError in pass_at_k when k > n. This can't go in __init__, because Metric.__call__ sets k/n with setattr afterwards.
2. Return 0.0 for k > n, matching GPassAtK.

I'd go with 1, since pass@k is undefined there and a silent 0.0 is also misleading. Happy to send a PR with a test for whichever you prefer.