fix: validate n >= 1 in the MMLU/GSM8K slice loaders (fixes #8)

src/steerbench/metrics.py

@@ -573,6 +573,8 @@ def load_mmlu_slice(
     cache miss. Slice size and seed are part of the cache key so different
     configs coexist.
     """
+    if n < 1:
+        raise ValueError(f"n must be >= 1, got {n}")
     path = _cache_path(cache_dir, f"mmlu_{n}_{seed}.json")
     if path.exists():
         rows = json.loads(path.read_text())
@@ -611,6 +613,8 @@ def load_gsm8k_slice(
     Gold answers are normalised from the dataset's ``#### <n>`` format at load
     time so scoring never re-parses them.
     """
+    if n < 1:
+        raise ValueError(f"n must be >= 1, got {n}")
     path = _cache_path(cache_dir, f"gsm8k_{n}_{seed}.json")
     if path.exists():
         rows = json.loads(path.read_text())

tests/test_metrics.py

@@ -24,6 +24,8 @@
     extract_mmlu_answer,
     formality_score,
     format_mmlu_prompt,
+    load_gsm8k_slice,
+    load_mmlu_slice,
     perplexity_from_token_nlls,
     repetition_rate,
     score_gsm8k,
@@ -220,6 +222,23 @@ def fake_generate(prompts: list[str]) -> list[str]:
     assert evaluate_gsm8k(examples, fake_generate) == 1.0
 
 
+# --------------------------------------------------------------------------- #
+# SLICE LOADERS — n validation (guard fires before any cache/datasets access)
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.parametrize("bad_n", [0, -3])
+def test_load_mmlu_slice_rejects_n_below_1(bad_n: int) -> None:
+    with pytest.raises(ValueError):
+        load_mmlu_slice(n=bad_n)
+
+
+@pytest.mark.parametrize("bad_n", [0, -3])
+def test_load_gsm8k_slice_rejects_n_below_1(bad_n: int) -> None:
+    with pytest.raises(ValueError):
+        load_gsm8k_slice(n=bad_n)
+
+
 def test_no_optional_deps_imported() -> None:
     # Importing steerbench.metrics must not pull in the heavy optional deps.
     # Checked in a fresh interpreter: a global sys.modules check is unreliable