Datasets:
File size: 7,849 Bytes
2b42456 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 | """Grading: canonical argument comparison and per-case verdicts."""
from __future__ import annotations
import math
import re
from dataclasses import dataclass
from typing import Any, Mapping, Sequence
from .parse import answer_text, parse_output
from .schema import Case, CaseResult, ExpectCall, ExpectNoCall, Protocol
from .tools import REGISTRY, canonical_name
@dataclass(frozen=True)
class ArgMatch:
ok: bool
missing: tuple[str, ...] = ()
mismatched: tuple[str, ...] = ()
extra: tuple[str, ...] = ()
def describe(self) -> str:
parts = []
if self.missing:
parts.append("missing=" + ",".join(self.missing))
if self.mismatched:
parts.append("mismatched=" + ",".join(self.mismatched))
if self.extra:
parts.append("extra=" + ",".join(self.extra))
return "; ".join(parts)
def _normalize(tool: str, key: str, value: Any) -> Any:
spec = REGISTRY.get(canonical_name(tool))
if spec is None:
return re.sub(r"\s+", " ", str(value)).strip().lower()
normalizer = spec.normalizers.get(key)
if normalizer is None:
return re.sub(r"\s+", " ", str(value)).strip().lower()
return normalizer(value)
def _equal(left: Any, right: Any) -> bool:
if isinstance(left, float) and isinstance(right, float):
return math.isclose(left, right, rel_tol=1e-6, abs_tol=1e-6)
if isinstance(left, (int, float)) and isinstance(right, (int, float)):
return math.isclose(float(left), float(right), rel_tol=1e-6, abs_tol=1e-6)
if isinstance(left, list) and isinstance(right, list):
return len(left) == len(right) and all(_equal(a, b) for a, b in zip(left, right))
if isinstance(left, Mapping) and isinstance(right, Mapping):
return set(left) == set(right) and all(_equal(left[k], right[k]) for k in left)
return left == right
def match_args(tool: str, expected: Mapping[str, Any], predicted: Mapping[str, Any]) -> ArgMatch:
missing: list[str] = []
mismatched: list[str] = []
for key, value in expected.items():
if key not in predicted:
missing.append(key)
continue
try:
left = _normalize(tool, key, value)
right = _normalize(tool, key, predicted[key])
except Exception:
left, right = value, predicted[key]
if not _equal(left, right):
mismatched.append(key)
spec = REGISTRY.get(canonical_name(tool))
allowed = set(spec.parameters) if spec else set(expected)
extra = tuple(sorted(key for key in predicted if key not in expected and key not in allowed))
optional_extra = tuple(sorted(key for key in predicted if key not in expected and key in allowed))
return ArgMatch(
ok=not missing and not mismatched,
missing=tuple(missing),
mismatched=tuple(mismatched),
extra=extra + optional_extra,
)
def _answer_matches(expect: ExpectNoCall, text: str) -> bool | None:
if not expect.answer_contains_any and not expect.answer_contains_all:
return None
haystack = text.lower()
if expect.answer_contains_all and not all(needle.lower() in haystack for needle in expect.answer_contains_all):
return False
if expect.answer_contains_any:
return any(any(needle.lower() in haystack for needle in group) for group in expect.answer_contains_any)
return True
def grade_case(case: Case, protocol: Protocol, raw_output: str) -> CaseResult:
parsed = parse_output(raw_output)
calls = parsed.calls
called = bool(calls)
text = answer_text(raw_output)
base = dict(
case_id=case.id,
protocol=protocol,
lang=case.lang,
category=case.category,
difficulty=case.difficulty,
malformed=parsed.malformed,
output=raw_output,
)
if isinstance(case.expect, ExpectCall):
if len(calls) != 1:
detail = "no call emitted" if not calls else f"{len(calls)} calls emitted (expected 1)"
return CaseResult(passed=False, called=called, expected_call=True, name_correct=False, args_exact=False, answer_ok=None, detail=detail, **base)
call = calls[0]
name_correct = call.name == canonical_name(case.expect.name)
args = match_args(case.expect.name, case.expect.args, call.args) if name_correct else ArgMatch(ok=False, mismatched=("name",))
detail = "" if (name_correct and args.ok) else f"predicted={call.canonical()}" + (f" ({args.describe()})" if args.describe() else "")
return CaseResult(
passed=bool(name_correct and args.ok),
called=True,
expected_call=True,
name_correct=name_correct,
args_exact=args.ok,
answer_ok=None,
detail=detail,
**base,
)
answer_ok = _answer_matches(case.expect, text)
passed = not calls and answer_ok is not False
detail = ""
if calls:
detail = "spurious call: " + "; ".join(c.canonical() for c in calls)
elif answer_ok is False:
detail = "answer failed constraint: " + text[:160]
return CaseResult(
passed=passed,
called=called,
expected_call=False,
name_correct=not calls,
args_exact=not calls,
answer_ok=answer_ok,
detail=detail,
**base,
)
def wilson_interval(successes: int, total: int, z: float = 1.96) -> tuple[float, float]:
if total == 0:
return (0.0, 0.0)
phat = successes / total
denominator = 1 + z * z / total
center = (phat + z * z / (2 * total)) / denominator
spread = z * math.sqrt(phat * (1 - phat) / total + z * z / (4 * total * total)) / denominator
return (max(0.0, center - spread), min(1.0, center + spread))
def rate(numerator: int, denominator: int) -> float:
return round(numerator / denominator, 4) if denominator else 0.0
def _group_summary(results: Sequence[CaseResult]) -> dict[str, Any]:
positives = [r for r in results if r.expected_call]
negatives = [r for r in results if not r.expected_call]
answered = [r for r in negatives if r.answer_ok is not None]
lo, hi = wilson_interval(sum(1 for r in results if r.passed), len(results))
return {
"cases": len(results),
"score": rate(sum(1 for r in results if r.passed), len(results)),
"score_ci95": [round(lo, 4), round(hi, 4)],
"correct_call_when_expected": rate(sum(1 for r in positives if r.passed), len(positives)),
"correct_abstention": rate(sum(1 for r in negatives if r.passed), len(negatives)),
"spurious_call_rate": rate(sum(1 for r in negatives if r.called), len(negatives)),
"name_accuracy": rate(sum(1 for r in positives if r.name_correct), len(positives)),
"args_accuracy": rate(sum(1 for r in positives if r.args_exact), len(positives)),
"answer_constraint_accuracy": rate(sum(1 for r in answered if r.answer_ok), len(answered)),
"malformed_rate": rate(sum(1 for r in results if r.malformed), len(results)),
}
def summarize(results: Sequence[CaseResult], model_id: str = "", protocol: Protocol | None = None) -> dict[str, Any]:
def slice_by(key: str) -> dict[str, Any]:
values: dict[str, list[CaseResult]] = {}
for result in results:
values.setdefault(getattr(result, key), []).append(result)
return {name: _group_summary(group) for name, group in sorted(values.items())}
return {
"benchmark": "ToolUse.1",
"model": model_id,
"protocol": protocol or "all",
"n_cases": len(results),
"overall": _group_summary(results),
"by_protocol": slice_by("protocol"),
"by_language": slice_by("lang"),
"by_category": slice_by("category"),
"by_difficulty": slice_by("difficulty"),
}
|