微调评估与交付：基线、过拟合、Adapter 合并和推理验收

import json


CASES = [
    ({"status": "offline", "temperature": 31}, "high"),
    ({"status": "warning", "temperature": 72}, "high"),
    ({"status": "online", "temperature": 42}, "low"),
]


def evaluate(predict) -> dict:
    valid_json = correct = 0
    details = []
    for inputs, expected in CASES:
        raw = predict(inputs)
        try:
            output = json.loads(raw)
            valid_json += 1
            ok = output.get("risk") == expected
        except (json.JSONDecodeError, AttributeError):
            ok = False
        correct += int(ok)
        details.append({"input": inputs, "raw": raw, "correct": ok})
    total = len(CASES)
    return {"accuracy": correct / total, "format_rate": valid_json / total, "details": details}


def baseline(inputs: dict) -> str:
    risk = "high" if inputs["status"] == "offline" or inputs["temperature"] >= 70 else "low"
    return json.dumps({"risk": risk})


print(evaluate(baseline))