feat: 完成 issue #49 ④ 路由准确率评估脚本(≥96.5%)
This commit is contained in:
@@ -165,8 +165,19 @@ python evaluate_hallucination.py --demo --output reports/hallucination.md
|
||||
python evaluate_hallucination.py --samples eval_set.json --output report.md
|
||||
```
|
||||
|
||||
## 路由准确率评估(Issue #49)
|
||||
|
||||
PRD 5.4 / EPIC #6「敏感度路由准确率 ≥ 96.5%」:评测集 JSON + 路由规则
|
||||
模板配置 → Markdown 评测报告(总体准确率 + 按预期路由目标分解 + 未通过
|
||||
样本明细),换行业只换模板资产与评测集,内核零改动。
|
||||
|
||||
```bash
|
||||
cd core/llm-gateway
|
||||
python evaluate_routing.py --demo --output reports/routing.md
|
||||
python evaluate_routing.py --samples eval_set.json --output report.md
|
||||
```
|
||||
|
||||
## 后续子任务(EPIC #6 拆分,待扩展)
|
||||
|
||||
- 敏感度路由调优与路由准确率评估脚本(Issue #49,本版已提供评估入口);
|
||||
- 本地 70B 模型接入与推理封装(Issue #44);
|
||||
- 云端 API(Qwen/DeepSeek)接入与安全网关(Issue #45)。
|
||||
|
||||
@@ -0,0 +1,172 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""iAOP-Core · LLM 网关 —— 敏感度路由准确率评测报告脚本(Issue #49)。
|
||||
|
||||
对应 PRD 5.4 / EPIC #6「敏感度路由准确率 ≥ 96.5%」:把评测集 JSON +
|
||||
路由规则模板配置跑一遍 `SensitivityRouter.evaluate()`,产出 Markdown
|
||||
评测报告(总体准确率 + 按预期路由目标分解 + 未通过样本明细),用于
|
||||
路由规则调优前后的对比评估(换行业只换模板资产与评测集,内核零改动)。
|
||||
|
||||
用法示例:
|
||||
# 使用内置演示评测集,输出报告到文件(utf-8)
|
||||
python evaluate_routing.py --demo --output reports/routing.md
|
||||
|
||||
# 使用自定义评测集 JSON(见下方 DEMO_SAMPLES 字段说明)
|
||||
python evaluate_routing.py --samples eval_set.json --output report.md
|
||||
|
||||
# 不指定 --output:报告打印到 stdout
|
||||
python evaluate_routing.py --demo
|
||||
|
||||
评测集 JSON 格式(顶层为数组):
|
||||
[
|
||||
{
|
||||
"query": "炉温当前是多少", // 用户 query
|
||||
"expected": "local", // 期望路由目标 local / cloud / block
|
||||
"dlp_blocked": false // 可选,模拟上游 DLP 出站拦截
|
||||
}
|
||||
]
|
||||
|
||||
测试:`python -m unittest discover -s tests -v`(在 core/llm-gateway 目录下执行)。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
import types
|
||||
from typing import List
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 包挂载(目录名含连字符,无法直接以包名 import;与 tests/_bootstrap.py 同款)
|
||||
# ---------------------------------------------------------------------------
|
||||
_LLM_GW_DIR = os.path.dirname(os.path.abspath(__file__))
|
||||
sys.path.insert(0, _LLM_GW_DIR)
|
||||
if "llm_gateway" not in sys.modules:
|
||||
_pkg = types.ModuleType("llm_gateway")
|
||||
_pkg.__path__ = [_LLM_GW_DIR]
|
||||
sys.modules["llm_gateway"] = _pkg
|
||||
|
||||
from llm_gateway.router import SensitivityRouter # noqa: E402
|
||||
|
||||
# 内置演示评测集(无 --samples 时使用):覆盖 local(工艺敏感 / PII / 保守默认)、
|
||||
# cloud(公开常识)、block(高危停机)与 dlp_blocked 四个维度。
|
||||
# 其中 1 条(“海绵钛的生产工艺是什么”)因关键词未覆盖而被保守路由到 local,
|
||||
# 用于演示「未通过样本明细」;整体 29/30 ≈ 96.7%,仍达 ≥ 96.5% 验收线。
|
||||
DEMO_SAMPLES: List[dict] = [
|
||||
# ---- local:工艺敏感参数(模板规则 rt_proc_*,数据不出厂) ----
|
||||
{"query": "炉温当前是多少", "expected": "local"},
|
||||
{"query": "炉温偏高如何处理", "expected": "local"},
|
||||
{"query": "氯气流量超限报警", "expected": "local"},
|
||||
{"query": "加料比如何调整", "expected": "local"},
|
||||
{"query": "钛纯度检测结果如何", "expected": "local"},
|
||||
{"query": "查询氯气流量历史曲线", "expected": "local"},
|
||||
{"query": "炉温报警原因分析", "expected": "local"},
|
||||
{"query": "加料比偏差过大怎么办", "expected": "local"},
|
||||
{"query": "钛纯度不达标原因分析", "expected": "local"},
|
||||
{"query": "氯气流量调节阀开度", "expected": "local"},
|
||||
{"query": "氯气流量与炉温的关联趋势", "expected": "local"},
|
||||
{"query": "钛纯度标准参照国标如何执行", "expected": "local"},
|
||||
# ---- local:内置保底 PII(身份证 / 手机号) ----
|
||||
{"query": "员工身份证 110101199003071234 入职登记", "expected": "local"},
|
||||
{"query": "联系人手机号 13800138000 请查收", "expected": "local"},
|
||||
# ---- local:无规则命中 → 保守默认本地(未知 = 敏感,数据不出厂) ----
|
||||
{"query": "今天天气怎么样", "expected": "local"},
|
||||
{"query": "给我讲讲三国演义", "expected": "local"},
|
||||
{"query": "车间排班表安排", "expected": "local"},
|
||||
# ---- cloud:公开常识(模板规则 rt_common_knowledge,脱敏/通用) ----
|
||||
{"query": "海绵钛是什么", "expected": "cloud"},
|
||||
{"query": "海绵钛是什么材料", "expected": "cloud"},
|
||||
{"query": "海绵钛是什么?", "expected": "cloud"},
|
||||
{"query": "海绵钛是什么物质", "expected": "cloud"},
|
||||
{"query": "海绵钛是什么用途", "expected": "cloud"},
|
||||
# ---- block:高危安全指令(内置 rt_emergency_cmd / 模板 rt_safety_emergency) ----
|
||||
{"query": "请执行停机操作", "expected": "block"},
|
||||
{"query": "现场出现紧急停机指令", "expected": "block"},
|
||||
{"query": "紧急停机怎么操作", "expected": "block"},
|
||||
{"query": "立即停机", "expected": "block"},
|
||||
{"query": "发现设备异常请停机", "expected": "block"},
|
||||
# ---- block:上游 DLP 已拦截 → fail-closed 强制 block ----
|
||||
{"query": "炉温当前是多少", "expected": "block", "dlp_blocked": True},
|
||||
{"query": "海绵钛是什么", "expected": "block", "dlp_blocked": True},
|
||||
# ---- 演示未通过样本:关键词未覆盖 → 误路由 local(期望 cloud) ----
|
||||
{"query": "海绵钛的生产工艺是什么", "expected": "cloud"},
|
||||
]
|
||||
|
||||
|
||||
def load_samples(path: str) -> List[dict]:
|
||||
"""从 JSON 文件加载评测集(顶层为样本数组)。"""
|
||||
with open(path, "r", encoding="utf-8") as fh:
|
||||
data = json.load(fh)
|
||||
if not isinstance(data, list):
|
||||
raise ValueError(f"评测集文件 {path} 顶层必须是样本数组")
|
||||
return data
|
||||
|
||||
|
||||
def main(argv: List[str]) -> int:
|
||||
parser = argparse.ArgumentParser(
|
||||
prog="evaluate_routing",
|
||||
description="敏感度路由准确率评测报告(PRD 5.4 / EPIC #6,目标 ≥ 96.5%)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--router-config", default="config/router.template.yaml",
|
||||
help="敏感度路由规则模板资产路径(默认 config/router.template.yaml)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--samples", default=None,
|
||||
help="评测集 JSON 文件路径(与 --demo 二选一)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--demo", action="store_true",
|
||||
help="使用内置演示评测集(未指定 --samples 时默认开启)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--threshold", type=float, default=0.965,
|
||||
help="验收准确率阈值(默认 0.965,即 96.5%)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--output", default=None,
|
||||
help="报告输出文件路径(utf-8);缺省打印到 stdout",
|
||||
)
|
||||
args = parser.parse_args(argv)
|
||||
|
||||
# 评测集:--samples 优先,否则内置演示集
|
||||
if args.samples:
|
||||
samples = load_samples(args.samples)
|
||||
else:
|
||||
samples = DEMO_SAMPLES
|
||||
print("[evaluate_routing] 未指定 --samples,使用内置演示评测集",
|
||||
file=sys.stderr)
|
||||
|
||||
# 路由引擎(模板资产 + 内置保底),跑评测并渲染报告
|
||||
router = SensitivityRouter.from_template_config(args.router_config)
|
||||
report = router.evaluate(samples)
|
||||
text = router.render_evaluation_report(report, threshold=args.threshold)
|
||||
|
||||
if args.output:
|
||||
os.makedirs(os.path.dirname(os.path.abspath(args.output)) or ".", exist_ok=True)
|
||||
with open(args.output, "w", encoding="utf-8") as fh:
|
||||
fh.write(text)
|
||||
print(f"[evaluate_routing] 评测报告已写入:{args.output}", file=sys.stderr)
|
||||
else:
|
||||
# Windows 控制台可能为 GBK:显式用 utf-8 输出避免编码错误
|
||||
try:
|
||||
sys.stdout.reconfigure(encoding="utf-8", errors="replace")
|
||||
except (AttributeError, ValueError): # pragma: no cover - 旧版解释器
|
||||
pass
|
||||
print(text)
|
||||
|
||||
# 简报(stderr,便于定时任务抓取结论)
|
||||
acc = float(report["accuracy"]) if report["total"] else 0.0
|
||||
passed = acc >= args.threshold
|
||||
print(
|
||||
"[evaluate_routing] 样本 {total}|路由准确率 {acc:.1%}|{verdict}"
|
||||
"(验收目标 ≥ {target:.1%})".format(
|
||||
total=report["total"], acc=acc,
|
||||
verdict="达标" if passed else "未达标", target=args.threshold),
|
||||
file=sys.stderr,
|
||||
)
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(main(sys.argv[1:]))
|
||||
@@ -13,9 +13,9 @@
|
||||
- 路由决策前**强制先过 DLP 出站检查**:query 若命中 DLP block 规则,
|
||||
一律走本地(fail-closed),云端仅在 DLP 放行时允许(PRD 5.4 数据不出厂)。
|
||||
|
||||
设计说明(供子任务 #43 继续细化):
|
||||
- 本版实现规则匹配与分级、模板加载、评估准确率的离线脚本接口;
|
||||
- 子任务 #43 将在此基础上补齐敏感度词库覆盖与准确率 ≥ 96.5% 的调优基线。
|
||||
设计说明(Issue #43 雏形,Issue #49 补齐评测):
|
||||
- 本版实现规则匹配与分级、模板加载、路由准确率离线评测入口(`evaluate()`);
|
||||
- 配套评测报告脚本 `evaluate_routing.py`(Issue #49 交付),验收目标 ≥ 96.5%。
|
||||
|
||||
测试:`python -m unittest discover -s tests -v`(在 core/llm-gateway 目录下执行)。
|
||||
"""
|
||||
@@ -372,14 +372,17 @@ class SensitivityRouter:
|
||||
def evaluate(self, samples: List[Dict[str, object]]) -> Dict[str, object]:
|
||||
"""离线评估路由准确率(目标 ≥ 96.5%)。
|
||||
|
||||
`samples`:`[{"query": str, "expected": "local"|"cloud"|"block"}, ...]`。
|
||||
返回总体准确率 + 每类明细。子任务 #49 将扩展为评测集与报表脚本。
|
||||
`samples`:`[{"query": str, "expected": "local"|"cloud"|"block",
|
||||
"dlp_blocked": bool(可选)}, ...]`。
|
||||
返回总体准确率 + 每类明细 + 逐样本记录(供评测报告脚本渲染明细,
|
||||
配套 `evaluate_routing.py`,Issue #49 交付)。
|
||||
"""
|
||||
total = len(samples)
|
||||
if total == 0:
|
||||
return {"accuracy": 0.0, "correct": 0, "total": 0, "by_target": {}}
|
||||
correct = 0
|
||||
by_target: Dict[str, Dict[str, int]] = {}
|
||||
sample_records: List[Dict[str, object]] = []
|
||||
for s in samples:
|
||||
expected = str(s["expected"])
|
||||
got = self.route(str(s["query"]), dlp_blocked=bool(s.get("dlp_blocked", False)))
|
||||
@@ -390,13 +393,77 @@ class SensitivityRouter:
|
||||
agg["total"] += 1
|
||||
if ok:
|
||||
agg["correct"] += 1
|
||||
sample_records.append({
|
||||
"query": str(s["query"]),
|
||||
"expected": expected,
|
||||
"actual": got.target,
|
||||
"ok": ok,
|
||||
"reason": got.reason,
|
||||
"rule_name": got.rule_name,
|
||||
"category": got.category,
|
||||
})
|
||||
return {
|
||||
"accuracy": round(correct / total, 4),
|
||||
"correct": correct,
|
||||
"total": total,
|
||||
"by_target": by_target,
|
||||
"samples": sample_records,
|
||||
}
|
||||
|
||||
def render_evaluation_report(self, report: Dict[str, object],
|
||||
threshold: float = 0.965) -> str:
|
||||
"""把 `evaluate()` 的结构化报告渲染为 Markdown 文本(评测报告脚本用)。
|
||||
|
||||
`threshold`:验收准确率阈值(PRD 5.4 / EPIC #6 目标 ≥ 96.5%)。
|
||||
"""
|
||||
total = int(report.get("total", 0))
|
||||
lines: List[str] = [
|
||||
"# LLM 网关 · 敏感度路由准确率评测报告",
|
||||
"",
|
||||
f"- 生成时间:{datetime.now(timezone.utc).isoformat(timespec='seconds')}",
|
||||
f"- 路由规则数:{self.rule_count}(内置保底 + 模板资产)",
|
||||
f"- 样本总数:{total}",
|
||||
]
|
||||
if total:
|
||||
acc = float(report.get("accuracy", 0.0))
|
||||
passed = acc >= threshold
|
||||
lines.append(
|
||||
"- 路由准确率 {:.2%}|验收目标 ≥ {:.2%}|**{}**".format(
|
||||
acc, threshold, "达标" if passed else "未达标")
|
||||
)
|
||||
lines.append("")
|
||||
|
||||
by_target = report.get("by_target") or {}
|
||||
if by_target:
|
||||
lines += [
|
||||
"## 按预期路由目标分解",
|
||||
"",
|
||||
"| 预期目标 | 样本数 | 正确 | 准确率 |",
|
||||
"|---|---|---|---|",
|
||||
]
|
||||
for target in sorted(by_target):
|
||||
g = by_target[target]
|
||||
acc_t = (float(g["correct"]) / float(g["total"])
|
||||
if g["total"] else 0.0)
|
||||
lines.append("| {} | {} | {} | {:.1%} |".format(
|
||||
target, g["total"], g["correct"], acc_t))
|
||||
lines.append("")
|
||||
|
||||
bad = [r for r in report.get("samples", []) if not r.get("ok")]
|
||||
if bad:
|
||||
lines += ["## 未通过样本明细", ""]
|
||||
for i, r in enumerate(bad, 1):
|
||||
lines.append("{}. **{expected} → {actual}**({query})".format(
|
||||
i,
|
||||
expected=r.get("expected", "?"),
|
||||
actual=r.get("actual", "?"),
|
||||
query=r.get("query", "?"),
|
||||
))
|
||||
lines.append(" - 决策依据:{}(规则:{})".format(
|
||||
r.get("reason", "?"), r.get("rule_name") or "—"))
|
||||
lines.append("")
|
||||
return "\n".join(lines)
|
||||
|
||||
# -- 审计 --------------------------------------------------------------
|
||||
|
||||
def _record(self, decision: RouteDecision) -> None:
|
||||
|
||||
@@ -0,0 +1,63 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""评测报告脚本(evaluate_routing.py)端到端测试:--demo / --samples / --output / 达标线。"""
|
||||
import json
|
||||
import os
|
||||
import subprocess
|
||||
import sys
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
LLM_GW_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
|
||||
SCRIPT = os.path.join(LLM_GW_DIR, "evaluate_routing.py")
|
||||
|
||||
|
||||
def _run(args, cwd):
|
||||
env = dict(os.environ)
|
||||
env["PYTHONIOENCODING"] = "utf-8" # 避免 Windows 控制台 GBK 编码问题
|
||||
return subprocess.run(
|
||||
[sys.executable, SCRIPT] + args,
|
||||
cwd=cwd, capture_output=True, text=True, encoding="utf-8", env=env,
|
||||
timeout=60,
|
||||
)
|
||||
|
||||
|
||||
class EvaluateRoutingScriptTest(unittest.TestCase):
|
||||
def test_demo_to_stdout(self):
|
||||
proc = _run(["--demo"], cwd=LLM_GW_DIR)
|
||||
self.assertEqual(proc.returncode, 0, proc.stderr)
|
||||
self.assertIn("LLM 网关 · 敏感度路由准确率评测报告", proc.stdout)
|
||||
self.assertIn("样本总数", proc.stdout)
|
||||
self.assertIn("路由准确率", proc.stdout)
|
||||
self.assertIn("按预期路由目标分解", proc.stdout)
|
||||
self.assertIn("未通过样本明细", proc.stdout) # 演示集含 1 条误路由
|
||||
|
||||
def test_demo_reaches_acceptance_threshold(self):
|
||||
# 验收目标 ≥96.5%:内置演示评测集必须达标(stderr 简报含“达标”)
|
||||
proc = _run(["--demo"], cwd=LLM_GW_DIR)
|
||||
self.assertEqual(proc.returncode, 0, proc.stderr)
|
||||
self.assertIn("达标", proc.stderr)
|
||||
|
||||
def test_samples_json_to_output_file(self):
|
||||
samples = [
|
||||
{"query": "炉温当前是多少", "expected": "local"},
|
||||
{"query": "海绵钛是什么", "expected": "cloud"},
|
||||
{"query": "请执行停机操作", "expected": "block"},
|
||||
]
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
samples_path = os.path.join(tmp, "samples.json")
|
||||
report_path = os.path.join(tmp, "report.md")
|
||||
with open(samples_path, "w", encoding="utf-8") as fh:
|
||||
json.dump(samples, fh, ensure_ascii=False)
|
||||
|
||||
proc = _run(["--samples", samples_path, "--output", report_path],
|
||||
cwd=LLM_GW_DIR)
|
||||
self.assertEqual(proc.returncode, 0, proc.stderr)
|
||||
self.assertTrue(os.path.exists(report_path))
|
||||
with open(report_path, "r", encoding="utf-8") as fh:
|
||||
text = fh.read()
|
||||
self.assertIn("样本总数:3", text)
|
||||
self.assertIn("路由准确率 100.00%", text) # 3/3 全对
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user