"""三池切分单元测试。 验证: - 三池互斥（question_id 无重叠） - test 池自然分布（correct_ratio=None） - save/load 往返一致 - 旧格式拒绝（无 test 键 → ValueError） - build_or_load_pools 冻结复用（pools.json 存在时不重切） """ from __future__ import annotations import json from typing import TYPE_CHECKING import pytest from app.harness.pools import ( build_pools, load_pools, save_pools, ) from core.types import GeneratedQuestion if TYPE_CHECKING: from pathlib import Path def _make_question(qid: str, task_type: str = "Action Reasoning") -> GeneratedQuestion: """构造测试用 GeneratedQuestion。 参数: qid: 题目 ID。 task_type: 题型。 返回: GeneratedQuestion 实例。 """ return GeneratedQuestion( question_id=qid, video_id="video_001", task_type=task_type, question=f"Question {qid}?", options=("A. opt1", "B. opt2", "C. opt3", "D. opt4"), answer="A", source_nodes=("node_1",), difficulty="medium", ) def _make_question_set( n: int, task_types: list[str] | None = None, ) -> list[GeneratedQuestion]: """构造 n 道题，交替分配题型。 参数: n: 题目数量。 task_types: 可选题型列表，轮转分配；None 默认 2 类。 返回: 题目列表。 """ types = task_types or ["Action Reasoning", "Scene Understanding"] return [_make_question(f"q_{i:04d}", types[i % len(types)]) for i in range(n)] def _make_correctness( questions: list[GeneratedQuestion], correct_ratio: float = 0.5, ) -> dict[str, bool]: """构造 correctness 字典，前 correct_ratio 比例标对。 参数: questions: 题目列表。 correct_ratio: 对题占比。 返回: question_id -> bool。 """ n_correct = round(len(questions) * correct_ratio) return {q.question_id: (i < n_correct) for i, q in enumerate(questions)} class TestBuildPoolsMutualExclusion: """三池 question_id 互斥验证。""" def test_build_pools_mutual_exclusion(self) -> None: """三池切分后，任意两池不共享 question_id。""" questions = _make_question_set(200) correctness = _make_correctness(questions, 0.5) pools = build_pools( questions, correctness, diag_cfg={ "size": 30, "correct_ratio": 0.5, "task_types": None, "seed": 42, "min_per_class": None, }, val_cfg={ "size": 30, "correct_ratio": 0.5, "task_types": None, "seed": 42, "min_per_class": None, }, test_cfg={"size": 30}, baseline_run_id="run_baseline", ) diag_ids = {q.question_id for q in pools.diagnosis} val_ids = {q.question_id for q in pools.validation} test_ids = {q.question_id for q in pools.test} assert diag_ids & val_ids == set(), "诊断池与验证池有重叠" assert diag_ids & test_ids == set(), "诊断池与测试池有重叠" assert val_ids & test_ids == set(), "验证池与测试池有重叠" assert len(diag_ids) == 30 assert len(val_ids) == 30 assert len(test_ids) == 30 class TestBuildPoolsTestNaturalDistribution: """test 池使用自然分布（correct_ratio=None）。""" def test_build_pools_test_natural_distribution(self) -> None: """test 池不强制对错比例，保留候选池的自然分布。 构造 correctness 为 50% 对/50% 错，diag/val 用 correct_ratio=0.3 强制裁剪，test 池走自然分布（correct_ratio=None）。验证 test 池 不受 correct_ratio 约束。 """ questions = _make_question_set(300) correctness = _make_correctness(questions, 0.5) pools = build_pools( questions, correctness, diag_cfg={ "size": 20, "correct_ratio": 0.3, "task_types": None, "seed": 42, "min_per_class": None, }, val_cfg={ "size": 20, "correct_ratio": 0.3, "task_types": None, "seed": 42, "min_per_class": None, }, test_cfg={"size": 20}, baseline_run_id="run_baseline", ) # diag/val 被 correct_ratio=0.3 裁剪：round(20*0.3) = 6 对, 14 错 diag_correct = sum(1 for q in pools.diagnosis if correctness[q.question_id]) val_correct = sum(1 for q in pools.validation if correctness[q.question_id]) assert diag_correct == 6, "诊断池应强制 30% 对题" assert val_correct == 6, "验证池应强制 30% 对题" # test 池自然分布：不受 correct_ratio 约束 assert len(pools.test) == 20 class TestSaveLoadPoolsRoundtrip: """save/load 往返一致验证。""" def test_save_load_pools_roundtrip(self, tmp_path: Path) -> None: """save_pools → load_pools 后全字段一致。""" questions = _make_question_set(100) correctness = _make_correctness(questions, 0.5) original = build_pools( questions, correctness, diag_cfg={ "size": 15, "correct_ratio": 0.5, "task_types": None, "seed": 42, "min_per_class": None, }, val_cfg={ "size": 15, "correct_ratio": 0.5, "task_types": None, "seed": 42, "min_per_class": None, }, test_cfg={"size": 15}, baseline_run_id="run_001", ) pools_path = tmp_path / "pools.json" save_pools(original, pools_path) restored = load_pools(pools_path) # 标量字段 assert restored.baseline_run_id == original.baseline_run_id assert restored.baseline_val_accuracy == pytest.approx(original.baseline_val_accuracy) assert restored.correctness == original.correctness # 三池逐题比对 for pool_name in ("diagnosis", "validation", "test"): orig_list = getattr(original, pool_name) rest_list = getattr(restored, pool_name) assert len(rest_list) == len(orig_list), f"{pool_name} 长度不一致" for o, r in zip(orig_list, rest_list, strict=False): assert o.question_id == r.question_id assert o.video_id == r.video_id assert o.task_type == r.task_type assert o.question == r.question assert o.options == r.options assert o.answer == r.answer assert o.source_nodes == r.source_nodes assert o.difficulty == r.difficulty class TestLoadPoolsOldFormatReject: """旧格式 pools.json（无 test 键）→ ValueError。""" def test_load_pools_old_format_reject(self, tmp_path: Path) -> None: """缺少 test 键的 pools.json 必须抛出 ValueError。""" old_format = { "baseline_run_id": "run_old", "baseline_val_accuracy": 0.5, "correctness": {}, "diagnosis": [], "validation": [], } pools_path = tmp_path / "pools.json" pools_path.write_text(json.dumps(old_format), encoding="utf-8") with pytest.raises(ValueError, match="旧格式"): load_pools(pools_path) class TestBuildOrLoadPoolsFrozen: """build_or_load_pools 冻结复用：pools.json 存在时原样加载不重切。""" def test_build_or_load_pools_frozen(self, tmp_path: Path) -> None: """pools.json 已存在时，build_or_load_pools 返回冻结内容。""" questions = _make_question_set(60) correctness = _make_correctness(questions, 0.5) frozen = build_pools( questions, correctness, diag_cfg={ "size": 10, "correct_ratio": 0.5, "task_types": None, "seed": 42, "min_per_class": None, }, val_cfg={ "size": 10, "correct_ratio": 0.5, "task_types": None, "seed": 42, "min_per_class": None, }, test_cfg={"size": 10}, baseline_run_id="run_frozen", ) pools_path = tmp_path / "pools.json" save_pools(frozen, pools_path) # build_or_load_pools 中 pools.json 存在 → 直接 load_pools # 此处直接测试 load_pools 行为等价 loaded = load_pools(pools_path) assert loaded.baseline_run_id == frozen.baseline_run_id assert loaded.baseline_val_accuracy == pytest.approx(frozen.baseline_val_accuracy) assert len(loaded.test) == len(frozen.test) assert len(loaded.validation) == len(frozen.validation) assert len(loaded.diagnosis) == len(frozen.diagnosis) # question_id 完全一致 for pool_name in ("diagnosis", "validation", "test"): orig_ids = [q.question_id for q in getattr(frozen, pool_name)] load_ids = [q.question_id for q in getattr(loaded, pool_name)] assert orig_ids == load_ids, f"{pool_name} 冻结后 ID 顺序不一致"