"""Smoke #9: the real HF datasets have the fields our loaders expect.
Requires network. Loads 3-row slices only."""
import pytest

pytestmark = [pytest.mark.network]


def _slice(dataset_id, split):
    from datasets import load_dataset
    return load_dataset(dataset_id, split=f"{split}[:3]")


def test_processbench_fields():
    ds = _slice("Qwen/ProcessBench", "gsm8k")
    row = ds[0]
    for k in ("problem", "steps", "label", "final_answer_correct"):
        assert k in row
    assert isinstance(row["steps"], list)
    assert isinstance(row["label"], int)


def test_prmbench_fields():
    ds = _slice("hitsmy/PRMBench_Preview", "train")
    row = ds[0]
    for k in ("original_process", "modified_process", "error_steps", "classification"):
        assert k in row


def test_math500_fields():
    ds = _slice("HuggingFaceH4/MATH-500", "test")
    row = ds[0]
    assert "problem" in row and "answer" in row


def test_aime_fields():
    ds = _slice("Maxwell-Jia/AIME_2024", "train")
    row = ds[0]
    assert "Problem" in row and "Answer" in row