fix(scoring): recognize TeX \over fractions in extract and normalize

pyproject.toml

@@ -18,7 +18,7 @@ dependencies = [
 ]
 
 [project.optional-dependencies]
-dev = ["pytest>=8.0", "ruff>=0.5", "mypy>=1.10"]
+dev = ["pytest>=8.0", "ruff>=0.5,<0.16", "mypy>=1.10"]
 
 [build-system]
 requires = ["hatchling"]

src/trinity/adapters/drop.py

@@ -130,9 +130,12 @@ def _maybe_unbox(segment: str) -> str:
     return boxed if boxed is not None else ""
 
 
-#: Surrounding punctuation stripped from a token, EXCLUDING the signs ``+``/``-`` — a
-#: leading sign is part of a number's value, not wrapping noise.
-_STRIP_EDGE = "".join(c for c in string.punctuation if c not in "+-")
+#: Surrounding punctuation stripped from a token, EXCLUDING the signs ``+``/``-`` and
+#: the decimal point ``.`` — a leading sign or decimal point is part of a number's
+#: value, not wrapping noise. A genuinely-trailing ``.`` (sentence period) is handled
+#: by the dedicated rstrip retry in :func:`_normalize_token`, which can tell it apart
+#: from a value-bearing leading point; a blanket edge-strip cannot.
+_STRIP_EDGE = "".join(c for c in string.punctuation if c not in "+-.")
 
 
 def _normalize_token(raw: str) -> str:
@@ -147,20 +150,27 @@ def _normalize_token(raw: str) -> str:
     dropped the ``-`` and left commas to break ``float()``) did not deliver.
 
     A token that is ALREADY a number is recognised before any punctuation is
-    stripped: the edge-strip set includes ``.``, so a leading-decimal token like
-    ``".5"`` would otherwise lose its point and normalize to ``"5.0"`` — equal to
-    a gold ``"5"`` (false positive) and unequal to the value-identical gold
-    ``"0.5"`` (false negative). The official DROP ``_remove_punc`` tests
-    ``_is_number`` first and leaves numbers untouched for exactly this reason."""
+    stripped: a leading-decimal token like ``".5"`` must not lose its point and
+    normalize to ``"5.0"`` — equal to a gold ``"5"`` (false positive) and unequal
+    to the value-identical gold ``"0.5"`` (false negative). The official DROP
+    ``_remove_punc`` tests ``_is_number`` first and leaves numbers untouched for
+    exactly this reason (issue #423). The float-first path alone only covers the
+    *bare* token: with ``.`` in the edge-strip set, wrapped forms like ``"$.5"``
+    and ``".5."`` still lost the leading point on the second-chance path. So the
+    edge strip excludes ``.`` entirely, and a genuinely-trailing period (``".5."``,
+    ``"16.."``) is retried with an explicit ``rstrip(".")`` — right-side dots are
+    sentence punctuation, left-side dots are value."""
     try:
         return str(float(raw.replace(",", "")))
     except ValueError:
         pass
     core = raw.strip(_STRIP_EDGE)
-    try:
-        return str(float(core.replace(",", "")))
-    except ValueError:
-        return _PUNCT.sub("", raw)
+    for cand in (core, core.rstrip(".")):
+        try:
+            return str(float(cand.replace(",", "")))
+        except ValueError:
+            continue
+    return _PUNCT.sub("", raw)
 
 
 def _split_internal_hyphens(token: str) -> list[str]:

src/trinity/orchestration/reward.py

@@ -505,6 +505,10 @@ def extract_last_number(text: str) -> str | None:
     candidates: list[tuple[int, str]] = []
     for _start, end, term in _iter_latex_frac_sqrt_spans(text):
         candidates.append((end, term))
+    # Classic TeX ``a\over b`` / ``{a\over b}`` — treat as one fraction term so the
+    # denominator digit is not taken alone (issue #484).
+    for m in re.finditer(r"\{[^{}]+\\over\s*[^{}]+\}|-?\d+\s*\\over\s*-?\d+", text):
+        candidates.append((m.end(), m.group(0)))
     # Match a simple fraction a/b FIRST (so "1/2" is kept whole, not read as "2"),
     # then decimals/integers like -1,234.56 or 42 or .5 ; require a digit somewhere.
     brace = r"\{(?:[^{}]|\{[^{}]*\})*\}"
@@ -521,7 +525,9 @@ def extract_last_number(text: str) -> str | None:
         candidates.append((m.end(), m.group(0).replace(",", "").replace(" ", "")))
     if not candidates:
         return None
-    candidates.sort(key=lambda item: item[0])
+    # Latest end wins; on a tie prefer the longer span so ``1\over 2`` beats the
+    # trailing digit ``2`` that ends at the same index (issue #484).
+    candidates.sort(key=lambda item: (item[0], len(item[1])))
     return candidates[-1][1]
 
 
@@ -820,6 +826,10 @@ def normalize_math_answer(ans: str | None) -> str:
     # The [dt]? family — \frac, \dfrac and \tfrac — all render the same value.
     s = _unwrap_latex_fractions(s)
     s = re.sub(r"\\[dt]?frac\s*(\d)\s*(\d)", r"\1/\2", s)
+    # Classic TeX ``\over`` fractions (``{1\over 2}``, ``1\over 2``) → slash form so
+    # they grade equal to ``1/2`` / ``\frac{1}{2}`` (issue #484).
+    s = re.sub(r"\{([^{}]+)\\over\s*([^{}]+)\}", r"(\1)/(\2)", s)
+    s = re.sub(r"(-?\d+(?:\.\d+)?)\\over\s*(-?\d+(?:\.\d+)?)", r"\1/\2", s)
     # \sqrt{x} -> sqrt(x), and the unbraced single-token \sqrt2 -> sqrt(2). Like the
     # \frac handling just above, the braced and bare forms render the SAME value, so
     # they must normalize identically (else \sqrt{2} is a false negative against a

tests/test_reward_tex_over.py

@@ -0,0 +1,26 @@
+"""Classic TeX ``\\over`` fractions must not grade as the denominator (issue #484)."""
+
+from trinity.orchestration.reward import extract_last_number, normalize_math_answer, score_text
+
+
+def test_extract_over_as_whole_term():
+    assert extract_last_number(r"1\over 2") == r"1\over 2"
+    assert extract_last_number(r"{1\over 2}") == r"{1\over 2}"
+    assert extract_last_number(r"answer is $1\over 2$") == r"1\over 2"
+
+
+def test_normalize_over_to_slash():
+    assert "\\over" not in normalize_math_answer(r"1\over 2")
+    assert "\\over" not in normalize_math_answer(r"{1\over 2}")
+    assert normalize_math_answer(r"1\over 2") == "1/2"
+
+
+def test_unboxed_over_grades_as_fraction_not_denominator():
+    assert score_text("math500", r"the answer is $1\over 2$", "1/2") == 1.0
+    assert score_text("math500", r"the answer is $1\over 2$", "2") == 0.0
+    assert score_text("math500", r"${1\over 2}$", r"\frac{1}{2}") == 1.0
+
+
+def test_boxed_over_equals_frac_gold():
+    assert score_text("math500", r"\boxed{1\over 2}", "1/2") == 1.0
+    assert score_text("math500", r"\boxed{{1\over 2}}", r"\frac{1}{2}") == 1.0