source/workbench/grading.py

"""Answer parsing shared by the task verifiers. Verifiers never trust the agent's format."""

import math
import re
from dataclasses import dataclass, field
from typing import Callable


@dataclass
class TaskInstance:
    instruction: str
    files: dict[str, str]
    grade: Callable[[str], float]
    oracle: str
    # Plausible but wrong answers, used only by tests.
    wrong: list[str] = field(default_factory=list)
    # Extra data for tests and workspace graders (never shown to the agent).
    meta: dict = field(default_factory=dict)


def norm(text: str) -> str:
    return " ".join(str(text).strip().strip("`'\"").lower().split())


def to_number(text: str):
    cleaned = re.sub(r"[\s,$€£%]", "", str(text)).strip("`'\"")
    cleaned = re.sub(r"(usd|eur|mm|mmol/l|g|s|sec|seconds)$", "", cleaned, flags=re.I)
    try:
        value = float(cleaned)
    except ValueError:
        return None
    return value if math.isfinite(value) else None


def number_close(text: str, target: float, abs_tol: float = 0.0, rel_tol: float = 0.0) -> bool:
    value = to_number(text)
    if value is None:
        return False
    return abs(value - target) <= max(abs_tol, rel_tol * abs(target)) + 1e-9


def split_parts(text: str, count: int, sep: str = ","):
    parts = [p.strip() for p in str(text).strip().strip("`'\"").split(sep)]
    return parts if len(parts) == count else None


def token_set(text: str):
    body = norm(text).strip("[](){}")
    if body == "":
        return None  # an empty submission never matches, even when the answer is "none"
    if body in ("none", "no", "nothing"):
        return frozenset()
    return frozenset(t.strip().strip("'\"") for t in re.split(r"[,\s;]+", body) if t.strip())


def parts_score(text: str, checks: list[Callable[[str], bool]], sep: str = ",") -> float:
    """Mean of per-part checks; 0 when the part count is wrong."""
    parts = split_parts(text, len(checks), sep)
    if parts is None and sep == ",":  # tolerate thousands separators such as 12,345.67
        parts = split_parts(re.sub(r"(?<=\d),(?=\d{3}(?!\d))", "", str(text)), len(checks), sep)
    if parts is None:
        return 0.0
    return sum(1.0 for part, check in zip(parts, checks) if check(part)) / len(checks)


def exact(target: str) -> Callable[[str], bool]:
    return lambda part: norm(part) == norm(target)


def close(target: float, abs_tol: float = 0.0, rel_tol: float = 0.0) -> Callable[[str], bool]:
    return lambda part: number_close(part, target, abs_tol, rel_tol)


def csv_text(header: list[str], rows: list[list]) -> str:
    lines = [",".join(header)]
    lines += [",".join(str(v) for v in row) for row in rows]
    return "\n".join(lines) + "\n"