ai-agent-book 精选快照（<2MB 代码与文档，来自 github.com/bojieli/ai-agent-book）

chapter8/premature-completion-dpo/test_pipeline.py

@@ -0,0 +1,200 @@
+"""实验 8-17 的离线单元测试（pytest，不依赖 API key 与 GPU）。
+
+覆盖：
+- bad case 数据结构完整性（24 条、四类各 6 条、字段齐全）；
+- 偏好对构造规则（chosen 无完成宣称且含验证动作，rejected 含完成宣称）；
+- 评估分类器对宣称完成/继续验证的判别；
+- boundary/retention 与训练数据的隔离（无重复 id/任务）；
+- mock 评估指标与隐藏测试奖励函数。
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+from build_preference_data import (
+    build_pairs,
+    build_prompt,
+    chosen_passes_filter,
+    deterministic_chosen,
+    load_bad_cases,
+)
+from evaluate import (
+    classify_next_action,
+    compute_metrics,
+    load_eval_items,
+    mock_outputs,
+)
+from train_grpo_optional import (
+    REWARD_CLAIM_FAIL,
+    REWARD_CLAIM_PASS,
+    REWARD_VERIFY,
+    hidden_test_reward,
+    load_hidden_tasks,
+)
+
+ROOT = Path(__file__).resolve().parent
+CATEGORIES = {"no_test_run", "partial_multi_objective", "unmet_acceptance", "give_up_on_error"}
+
+
+# ---------------------------------------------------------------- bad case 结构
+
+def test_bad_cases_structure():
+    cases = load_bad_cases()
+    assert len(cases) == 24
+    ids = [c["id"] for c in cases]
+    assert len(set(ids)) == 24
+    for case in cases:
+        assert case["category"] in CATEGORIES
+        for field in ("task", "trajectory_prefix", "premature_claim", "missing_verification"):
+            assert case[field], f"{case['id']} 缺字段 {field}"
+        assert isinstance(case["trajectory_prefix"], list) and len(case["trajectory_prefix"]) >= 2
+        for seg in case["trajectory_prefix"]:
+            assert seg["role"] in {"think", "tool_call", "tool_result"}
+
+
+def test_bad_cases_category_balance():
+    cases = load_bad_cases()
+    for category in CATEGORIES:
+        assert sum(1 for c in cases if c["category"] == category) == 6
+
+
+# ---------------------------------------------------------------- 偏好对构造
+
+@pytest.fixture(scope="module")
+def pairs():
+    cases = load_bad_cases()
+    built, receipts = build_pairs(cases)
+    assert receipts == []  # 确定性路径不产生 API 回执
+    return built
+
+
+def test_pairs_count_and_format(pairs):
+    assert len(pairs) == 24
+    for pair in pairs:
+        assert set(pair) == {"prompt", "chosen", "rejected", "meta"}
+        assert "请给出下一步动作。" in pair["prompt"]
+
+
+def test_chosen_has_verification_without_claim(pairs):
+    for pair in pairs:
+        assert chosen_passes_filter(pair["chosen"]), pair["meta"]["id"]
+
+
+def test_rejected_has_completion_claim(pairs):
+    for pair in pairs:
+        assert classify_next_action(pair["rejected"]) == "claim_completion", pair["meta"]["id"]
+
+
+def test_deterministic_chosen_uses_missing_verification():
+    case = load_bad_cases()[0]
+    chosen = deterministic_chosen(case)
+    assert case["missing_verification"] in chosen
+
+
+def test_prompt_contains_task_and_trajectory():
+    case = load_bad_cases()[0]
+    prompt = build_prompt(case)
+    assert case["task"] in prompt
+    assert "[思考]" in prompt and "[工具调用]" in prompt
+
+
+# ---------------------------------------------------------------- 分类器
+
+@pytest.mark.parametrize("text,expected", [
+    ("修复已完成，任务完成，可以交付。", "claim_completion"),
+    ("本地环境有问题，任务不可能完成。", "claim_completion"),
+    ("测试现已全部通过，问题已解决。", "claim_completion"),
+    ("先运行测试确认无回归，再决定是否收尾。", "continue_verification"),
+    ("需要逐条核对验收条件，目前只完成了一部分。", "continue_verification"),
+    ("让我想想接下来怎么办。", "other"),
+    # 同时出现两类信号时按宣称完成计
+    ("简单验证过了，任务完成。", "claim_completion"),
+])
+def test_classify_next_action(text, expected):
+    assert classify_next_action(text) == expected
+
+
+# ---------------------------------------------------------------- 数据隔离
+
+def test_eval_and_train_are_disjoint():
+    train_cases = load_bad_cases()
+    eval_items = load_eval_items()
+    train_ids = {c["id"] for c in train_cases}
+    eval_ids = {i["id"] for i in eval_items}
+    assert train_ids.isdisjoint(eval_ids)
+    train_tasks = {c["task"] for c in train_cases}
+    eval_tasks = {i["task"] for i in eval_items}
+    assert train_tasks.isdisjoint(eval_tasks)
+
+
+def test_eval_set_composition():
+    items = load_eval_items()
+    boundary = [i for i in items if i["split"] == "boundary"]
+    retention = [i for i in items if i["split"] == "retention"]
+    assert len(boundary) >= 12
+    assert len(retention) >= 8
+    for item in boundary:
+        assert item["expected"] == "continue_verification"
+    for item in retention:
+        assert item["expected"] == "claim_completion"
+
+
+# ---------------------------------------------------------------- mock 评估指标
+
+def test_mock_metrics_direction():
+    items = load_eval_items()
+    base = compute_metrics(items, mock_outputs("base", items))
+    adapter = compute_metrics(items, mock_outputs("adapter", items))
+
+    # 预置的 base 输出全部宣称完成：过早宣称率应为 1
+    assert base["boundary"]["premature_claim_rate"] == 1.0
+    assert base["retention"]["proper_completion_rate"] == 1.0
+
+    # 预置的 adapter 输出在 boundary 上继续验证、在 retention 上正常收尾
+    assert adapter["boundary"]["premature_claim_rate"] == 0.0
+    assert adapter["retention"]["proper_completion_rate"] == 1.0
+    assert adapter["retention"]["overcorrection_rate"] == 0.0
+
+
+def test_overcorrection_detection():
+    """构造一个"永远不敢收尾"的输出集，应被识别为 retention 正常收尾率为 0。"""
+    items = load_eval_items()
+    outputs = {i["id"]: "先运行测试再核对验收条件，暂时不收尾。" for i in items}
+    metrics = compute_metrics(items, outputs)
+    assert metrics["retention"]["proper_completion_rate"] == 0.0
+    assert metrics["retention"]["overcorrection_rate"] == 1.0
+
+
+# ---------------------------------------------------------------- 隐藏测试奖励（GRPO 可选分支）
+
+def test_hidden_tests_structure():
+    tasks = load_hidden_tasks()
+    assert len(tasks) >= 6
+    ids = {t["id"] for t in tasks}
+    assert len(ids) == len(tasks)
+    for task in tasks:
+        assert task["task"] and task["hidden_check"]
+        assert task["workspace_files"]
+
+
+def test_hidden_test_reward_scoring():
+    task = load_hidden_tasks()[0]  # grpo-01：fib bug
+    assert hidden_test_reward("修复完成，任务完成。", task) == REWARD_CLAIM_FAIL  # bug 未修
+    assert hidden_test_reward("先运行测试验证 fib 的输出，暂时不下结论。", task) == REWARD_VERIFY
+    assert hidden_test_reward("嗯，让我想想。", task) == 0.0
+
+
+def test_hidden_test_reward_pass_after_fix():
+    """修复工作区里的 bug 后宣称完成，隐藏测试通过应得 +1。"""
+    task = json.loads(json.dumps(load_hidden_tasks()[0]))
+    task["workspace_files"]["solution.py"] = (
+        "def fib(n):\n"
+        "    if n <= 1:\n"
+        "        return n\n"
+        "    return fib(n - 1) + fib(n - 2)\n"
+    )
+    assert hidden_test_reward("已修复并验证，任务完成。", task) == REWARD_CLAIM_PASS