squad

Metric SQuAD from torchmetrics (torchmetrics.SQuAD)

When to invoke this skill

The user has predictions + ground truth and asks to evaluate with SQuAD, or mentions torchmetrics.SQuAD directly, or wants the standard torchmetrics implementation.

Reference signature

from torchmetrics import SQuAD

# _SQuAD(**kwargs: Any) -> None

Library docstring

Wrapper for deprecated import.

>>> preds = [{"prediction_text": "1976", "id": "56e10a3be3433e1400422b22"}]
>>> target = [{"answers": {"answer_start": [97], "text": ["1976"]}, "id": "56e10a3be3433e1400422b22"}]
>>> squad = _SQuAD()
>>> squad(preds, target)
{'exact_match': tensor(100.), 'f1': tensor(100.)}

Quick recipe

import torchmetrics as _m
score = _m.SQuAD(y_true, y_pred)

Don'ts

• Don't reimplement when the library version handles edge cases (NaN, ties, empty inputs) better than a hand-rolled formula.
• Always check the library version's argument order — sklearn is (y_true, y_pred) while torchmetrics is (preds, target).