feat(question_gen): add lightweight 4-gate quality check
Implement 4 concurrent LLM-based quality gates for generated questions: - key_verify: validates answer evidence in source material - blind_answer: rejects questions answerable without video context - multi_true: detects ambiguous multi-correct options - leak_test: per-family shortcut detection (5 probe templates) Includes run_gates orchestrator with verbatim_ratio short-circuit, JSON response parsing with fallback, and 9 unit tests (all passing). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,471 @@
|
||||
"""轻量四门质量检查 — 对生成题目执行 4 个独立的 LLM 质量门控。
|
||||
|
||||
四门：
|
||||
1. key_verify: 验证答案在来源素材中有证据支撑。
|
||||
2. blind_answer: 无上下文时 LLM 能否答对（若答对 → 题目泄漏）。
|
||||
3. multi_true: 检测是否有多个选项可被视为正确。
|
||||
4. leak_test: 按家族特定模板探测答题捷径。
|
||||
|
||||
设计要点：
|
||||
- run_gates 先做 verbatim_ratio 前置短路（> 0.5 直接 FAIL key_verify）。
|
||||
- 四门并发执行（asyncio.gather）。
|
||||
- 每门加载 store/prompts/question_gen/ 下对应模板，构造 messages，调 LLM。
|
||||
- LLM 返回 JSON {"verdict": "pass"|"fail", "reason": "..."}，解析失败视为 FAIL。
|
||||
- session_id 必须透传至每次 LLM 调用（遥测关联）。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import json
|
||||
from dataclasses import dataclass, field
|
||||
from enum import Enum
|
||||
from pathlib import Path
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from loguru import logger
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from app.question_gen.families import QuestionFamilySpec
|
||||
from app.question_gen.postprocess import PostprocessResult
|
||||
from app.tree.index import TreeIndex
|
||||
from core.protocols import LLMProvider
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 常量
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
_PROMPTS_DIR = Path(__file__).resolve().parent.parent.parent / "store" / "prompts" / "question_gen"
|
||||
|
||||
_VERBATIM_THRESHOLD = 0.5
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# CandidateQuestion（Task 5 尚未实现，本地定义，后续迁移至 generator_v2.py）
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class CandidateQuestion:
|
||||
"""候选题目 — 出题管线生成、待门控审核的题目数据。
|
||||
|
||||
属性:
|
||||
question_id: 题目唯一标识。
|
||||
video_id: 所属视频标识。
|
||||
task_type: 题型（如 "Action Reasoning"）。
|
||||
skill_target: 目标失败机制编号（M1-M5）。
|
||||
question: 题目文本。
|
||||
options: 选项元组（如 ("A. ...", "B. ...", "C. ...", "D. ...")）。
|
||||
answer: 正确答案字母（如 "A"）。
|
||||
source_nodes: 来源节点 ID 元组。
|
||||
difficulty: 难度等级。
|
||||
subtitle_sentences: 字幕句子元组。
|
||||
frame_paths: 帧图片路径元组。
|
||||
"""
|
||||
|
||||
question_id: str
|
||||
video_id: str
|
||||
task_type: str
|
||||
skill_target: str
|
||||
question: str
|
||||
options: tuple[str, ...]
|
||||
answer: str
|
||||
source_nodes: tuple[str, ...]
|
||||
difficulty: str
|
||||
subtitle_sentences: tuple[str, ...] = field(default_factory=tuple)
|
||||
frame_paths: tuple[str, ...] = field(default_factory=tuple)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 门控结果类型
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class GateVerdict(Enum):
|
||||
"""门控判决枚举。"""
|
||||
|
||||
PASS = "pass"
|
||||
FAIL = "fail"
|
||||
SKIP = "skip"
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class GateResult:
|
||||
"""单门判决结果。
|
||||
|
||||
属性:
|
||||
verdict: 门控判决（PASS / FAIL / SKIP）。
|
||||
reason: 判决原因说明。
|
||||
raw_response: LLM 原始返回内容。
|
||||
"""
|
||||
|
||||
verdict: GateVerdict
|
||||
reason: str
|
||||
raw_response: str
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class GateReport:
|
||||
"""四门判决汇总报告。
|
||||
|
||||
属性:
|
||||
key_verify: 关键验证门结果。
|
||||
blind_answer: 盲答门结果。
|
||||
multi_true: 多正确门结果。
|
||||
leak_test: 泄漏测试门结果。
|
||||
"""
|
||||
|
||||
key_verify: GateResult
|
||||
blind_answer: GateResult
|
||||
multi_true: GateResult
|
||||
leak_test: GateResult
|
||||
|
||||
@property
|
||||
def passed(self) -> bool:
|
||||
"""全门 PASS 或 SKIP 时视为通过。"""
|
||||
for gate in (self.key_verify, self.blind_answer, self.multi_true, self.leak_test):
|
||||
if gate.verdict == GateVerdict.FAIL:
|
||||
return False
|
||||
return True
|
||||
|
||||
@property
|
||||
def reject_reason(self) -> str | None:
|
||||
"""首个 FAIL 门的 reason，全部通过返回 None。"""
|
||||
for name, gate in [
|
||||
("key_verify", self.key_verify),
|
||||
("blind_answer", self.blind_answer),
|
||||
("multi_true", self.multi_true),
|
||||
("leak_test", self.leak_test),
|
||||
]:
|
||||
if gate.verdict == GateVerdict.FAIL:
|
||||
return f"[{name}] {gate.reason}"
|
||||
return None
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 内部工具函数
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _load_prompt_template(filename: str) -> str:
|
||||
"""加载 store/prompts/question_gen/ 下的模板文件。
|
||||
|
||||
参数:
|
||||
filename: 模板文件名（如 "gate_key_verify.md"）。
|
||||
|
||||
返回:
|
||||
模板内容字符串。
|
||||
|
||||
异常:
|
||||
FileNotFoundError: 模板文件不存在。
|
||||
"""
|
||||
path = _PROMPTS_DIR / filename
|
||||
if not path.exists():
|
||||
msg = f"门控模板文件不存在: {path}"
|
||||
raise FileNotFoundError(msg)
|
||||
return path.read_text(encoding="utf-8")
|
||||
|
||||
|
||||
def _resolve_source_text(candidate: CandidateQuestion, tree: TreeIndex) -> str:
|
||||
"""从树中解析候选题的来源节点文本。
|
||||
|
||||
遍历树的所有层级，匹配 source_nodes 中的 ID，提取对应描述文本。
|
||||
|
||||
参数:
|
||||
candidate: 候选题目。
|
||||
tree: 视频树索引。
|
||||
|
||||
返回:
|
||||
拼接的来源节点描述文本。
|
||||
"""
|
||||
texts: list[str] = []
|
||||
target_ids = set(candidate.source_nodes)
|
||||
|
||||
for l1 in tree.roots:
|
||||
if l1.id in target_ids:
|
||||
texts.append(f"[L1 {l1.id}] {l1.card.scene_summary}")
|
||||
for l2 in l1.children:
|
||||
if l2.id in target_ids:
|
||||
texts.append(f"[L2 {l2.id}] {l2.card.event_description}")
|
||||
for l3 in l2.children:
|
||||
if l3.id in target_ids:
|
||||
texts.append(f"[L3 {l3.id}] {l3.card.frame_summary}")
|
||||
|
||||
if not texts:
|
||||
logger.warning(
|
||||
"未找到来源节点: candidate={}, source_nodes={}",
|
||||
candidate.question_id,
|
||||
candidate.source_nodes,
|
||||
)
|
||||
return "(no source material found)"
|
||||
|
||||
return "\n".join(texts)
|
||||
|
||||
|
||||
def _format_options(options: tuple[str, ...]) -> str:
|
||||
"""将选项元组格式化为可读字符串。"""
|
||||
return "\n".join(options)
|
||||
|
||||
|
||||
def _parse_gate_response(raw_content: str) -> tuple[GateVerdict, str]:
|
||||
"""解析 LLM 返回的门控 JSON 响应。
|
||||
|
||||
期望格式: {"verdict": "pass"|"fail", "reason": "..."}
|
||||
解析失败时返回 FAIL + parse_error。
|
||||
|
||||
参数:
|
||||
raw_content: LLM 返回的原始文本。
|
||||
|
||||
返回:
|
||||
(verdict, reason) 元组。
|
||||
"""
|
||||
try:
|
||||
# 尝试从文本中提取 JSON（可能被 markdown 代码块包裹）
|
||||
content = raw_content.strip()
|
||||
if "```" in content:
|
||||
# 提取代码块中的内容
|
||||
parts = content.split("```")
|
||||
for part in parts:
|
||||
stripped = part.strip()
|
||||
if stripped.startswith("json"):
|
||||
stripped = stripped[4:].strip()
|
||||
if stripped.startswith("{"):
|
||||
content = stripped
|
||||
break
|
||||
|
||||
data = json.loads(content)
|
||||
verdict_str = data.get("verdict", "").lower()
|
||||
reason = data.get("reason", "")
|
||||
|
||||
if verdict_str == "pass":
|
||||
return GateVerdict.PASS, reason
|
||||
elif verdict_str == "fail":
|
||||
return GateVerdict.FAIL, reason
|
||||
else:
|
||||
return GateVerdict.FAIL, f"parse_error: invalid verdict '{verdict_str}'"
|
||||
|
||||
except (json.JSONDecodeError, AttributeError, TypeError) as e:
|
||||
return GateVerdict.FAIL, f"parse_error: {e}"
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 四门实现
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
async def _gate_key_verify(
|
||||
candidate: CandidateQuestion,
|
||||
tree: TreeIndex,
|
||||
llm: LLMProvider,
|
||||
*,
|
||||
session_id: str,
|
||||
) -> GateResult:
|
||||
"""关键验证门 — 检查答案在来源素材中是否有证据支撑。
|
||||
|
||||
参数:
|
||||
candidate: 候选题目。
|
||||
tree: 视频树索引。
|
||||
llm: LLM 调用端口。
|
||||
session_id: 会话 ID（遥测关联）。
|
||||
|
||||
返回:
|
||||
GateResult 实例。
|
||||
"""
|
||||
source_text = _resolve_source_text(candidate, tree)
|
||||
template = _load_prompt_template("gate_key_verify.md")
|
||||
prompt = template.format(
|
||||
source_text=source_text,
|
||||
question=candidate.question,
|
||||
options=_format_options(candidate.options),
|
||||
answer=candidate.answer,
|
||||
)
|
||||
|
||||
response = await llm.chat(
|
||||
[{"role": "user", "content": prompt}],
|
||||
session_id=session_id,
|
||||
)
|
||||
|
||||
verdict, reason = _parse_gate_response(response.content)
|
||||
return GateResult(verdict=verdict, reason=reason, raw_response=response.content)
|
||||
|
||||
|
||||
async def _gate_blind_answer(
|
||||
candidate: CandidateQuestion,
|
||||
llm: LLMProvider,
|
||||
*,
|
||||
session_id: str,
|
||||
) -> GateResult:
|
||||
"""盲答门 — 无上下文时 LLM 能否答对（答对 → 题目泄漏）。
|
||||
|
||||
参数:
|
||||
candidate: 候选题目。
|
||||
llm: LLM 调用端口。
|
||||
session_id: 会话 ID（遥测关联）。
|
||||
|
||||
返回:
|
||||
GateResult 实例。
|
||||
"""
|
||||
template = _load_prompt_template("gate_blind_answer.md")
|
||||
prompt = template.format(
|
||||
question=candidate.question,
|
||||
options=_format_options(candidate.options),
|
||||
)
|
||||
|
||||
response = await llm.chat(
|
||||
[{"role": "user", "content": prompt}],
|
||||
session_id=session_id,
|
||||
)
|
||||
|
||||
verdict, reason = _parse_gate_response(response.content)
|
||||
return GateResult(verdict=verdict, reason=reason, raw_response=response.content)
|
||||
|
||||
|
||||
async def _gate_multi_true(
|
||||
candidate: CandidateQuestion,
|
||||
tree: TreeIndex,
|
||||
llm: LLMProvider,
|
||||
*,
|
||||
session_id: str,
|
||||
) -> GateResult:
|
||||
"""多正确门 — 检测是否有多个选项可被视为正确。
|
||||
|
||||
参数:
|
||||
candidate: 候选题目。
|
||||
tree: 视频树索引。
|
||||
llm: LLM 调用端口。
|
||||
session_id: 会话 ID（遥测关联）。
|
||||
|
||||
返回:
|
||||
GateResult 实例。
|
||||
"""
|
||||
source_text = _resolve_source_text(candidate, tree)
|
||||
template = _load_prompt_template("gate_multi_true.md")
|
||||
prompt = template.format(
|
||||
source_text=source_text,
|
||||
question=candidate.question,
|
||||
options=_format_options(candidate.options),
|
||||
)
|
||||
|
||||
response = await llm.chat(
|
||||
[{"role": "user", "content": prompt}],
|
||||
session_id=session_id,
|
||||
)
|
||||
|
||||
verdict, reason = _parse_gate_response(response.content)
|
||||
return GateResult(verdict=verdict, reason=reason, raw_response=response.content)
|
||||
|
||||
|
||||
async def _gate_leak_test(
|
||||
candidate: CandidateQuestion,
|
||||
family_spec: QuestionFamilySpec,
|
||||
llm: LLMProvider,
|
||||
*,
|
||||
session_id: str,
|
||||
) -> GateResult:
|
||||
"""泄漏测试门 — 按家族特定模板探测答题捷径。
|
||||
|
||||
参数:
|
||||
candidate: 候选题目。
|
||||
family_spec: 问题家族规格（含 leak_profile）。
|
||||
llm: LLM 调用端口。
|
||||
session_id: 会话 ID（遥测关联）。
|
||||
|
||||
返回:
|
||||
GateResult 实例。
|
||||
"""
|
||||
probe_template_name = family_spec.leak_profile.probe_template
|
||||
template = _load_prompt_template(probe_template_name)
|
||||
prompt = template.format(
|
||||
question=candidate.question,
|
||||
options=_format_options(candidate.options),
|
||||
answer=candidate.answer,
|
||||
)
|
||||
|
||||
response = await llm.chat(
|
||||
[{"role": "user", "content": prompt}],
|
||||
session_id=session_id,
|
||||
)
|
||||
|
||||
verdict, reason = _parse_gate_response(response.content)
|
||||
return GateResult(verdict=verdict, reason=reason, raw_response=response.content)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 编排入口
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
async def run_gates(
|
||||
candidate: CandidateQuestion,
|
||||
tree: TreeIndex,
|
||||
llm: LLMProvider,
|
||||
family_spec: QuestionFamilySpec,
|
||||
postprocess: PostprocessResult,
|
||||
*,
|
||||
session_id: str,
|
||||
) -> GateReport:
|
||||
"""编排四门并发执行，返回汇总报告。
|
||||
|
||||
前置规则：verbatim_ratio > 0.5 时直接短路 key_verify 为 FAIL（不调 LLM）。
|
||||
|
||||
参数:
|
||||
candidate: 候选题目。
|
||||
tree: 视频树索引。
|
||||
llm: LLM 调用端口。
|
||||
family_spec: 问题家族规格。
|
||||
postprocess: 后处理结果（含 verbatim_ratio）。
|
||||
session_id: 会话 ID（遥测关联）。
|
||||
|
||||
返回:
|
||||
GateReport 四门汇总。
|
||||
"""
|
||||
# Phase 1: verbatim 前置短路
|
||||
if postprocess.verbatim_ratio > _VERBATIM_THRESHOLD:
|
||||
logger.info(
|
||||
"verbatim_ratio={:.3f} > {:.1f}，key_verify 短路 FAIL",
|
||||
postprocess.verbatim_ratio,
|
||||
_VERBATIM_THRESHOLD,
|
||||
)
|
||||
key_verify_result = GateResult(
|
||||
verdict=GateVerdict.FAIL,
|
||||
reason=f"verbatim_ratio={postprocess.verbatim_ratio:.3f} exceeds threshold {_VERBATIM_THRESHOLD}",
|
||||
raw_response="",
|
||||
)
|
||||
# 其余三门仍然并发执行（收集完整诊断信息）
|
||||
blind_result, multi_result, leak_result = await asyncio.gather(
|
||||
_gate_blind_answer(candidate, llm, session_id=session_id),
|
||||
_gate_multi_true(candidate, tree, llm, session_id=session_id),
|
||||
_gate_leak_test(candidate, family_spec, llm, session_id=session_id),
|
||||
)
|
||||
return GateReport(
|
||||
key_verify=key_verify_result,
|
||||
blind_answer=blind_result,
|
||||
multi_true=multi_result,
|
||||
leak_test=leak_result,
|
||||
)
|
||||
|
||||
# Phase 2: 四门并发执行
|
||||
key_result, blind_result, multi_result, leak_result = await asyncio.gather(
|
||||
_gate_key_verify(candidate, tree, llm, session_id=session_id),
|
||||
_gate_blind_answer(candidate, llm, session_id=session_id),
|
||||
_gate_multi_true(candidate, tree, llm, session_id=session_id),
|
||||
_gate_leak_test(candidate, family_spec, llm, session_id=session_id),
|
||||
)
|
||||
|
||||
report = GateReport(
|
||||
key_verify=key_result,
|
||||
blind_answer=blind_result,
|
||||
multi_true=multi_result,
|
||||
leak_test=leak_result,
|
||||
)
|
||||
|
||||
if report.passed:
|
||||
logger.debug("四门全部通过: question_id={}", candidate.question_id)
|
||||
else:
|
||||
logger.info(
|
||||
"门控拒绝: question_id={}, reason={}",
|
||||
candidate.question_id,
|
||||
report.reject_reason,
|
||||
)
|
||||
|
||||
return report
|
||||
@@ -0,0 +1,24 @@
|
||||
# Blind Answer Gate
|
||||
|
||||
You are evaluating whether a multiple-choice question can be answered correctly **without any video context**, using only common sense or option patterns.
|
||||
|
||||
## Question
|
||||
|
||||
{question}
|
||||
|
||||
## Options
|
||||
|
||||
{options}
|
||||
|
||||
## Instructions
|
||||
|
||||
1. Read the question and options WITHOUT any source material.
|
||||
2. Try to determine the correct answer using only common sense, option length patterns, or grammatical cues.
|
||||
3. If you can confidently pick the correct answer → verdict "fail" (the question leaks information).
|
||||
4. If you cannot determine the answer without context → verdict "pass" (the question genuinely requires video understanding).
|
||||
|
||||
## Response Format (strict JSON)
|
||||
|
||||
```json
|
||||
{{"verdict": "pass" or "fail", "reason": "brief explanation"}}
|
||||
```
|
||||
@@ -0,0 +1,35 @@
|
||||
# Key Verify Gate
|
||||
|
||||
You are a quality-control judge for video understanding questions.
|
||||
|
||||
## Task
|
||||
|
||||
Given the source material from a video and a multiple-choice question with its designated correct answer, determine whether the correct answer is **supported by evidence** in the source material.
|
||||
|
||||
## Source Material
|
||||
|
||||
{source_text}
|
||||
|
||||
## Question
|
||||
|
||||
{question}
|
||||
|
||||
## Options
|
||||
|
||||
{options}
|
||||
|
||||
## Designated Correct Answer
|
||||
|
||||
{answer}
|
||||
|
||||
## Instructions
|
||||
|
||||
1. Read the source material carefully.
|
||||
2. Determine if the designated correct answer can be derived or inferred from the source material.
|
||||
3. If evidence supports the answer, verdict is "pass". If not, verdict is "fail".
|
||||
|
||||
## Response Format (strict JSON)
|
||||
|
||||
```json
|
||||
{{"verdict": "pass" or "fail", "reason": "brief explanation"}}
|
||||
```
|
||||
@@ -0,0 +1,32 @@
|
||||
# Leak Test: Option Length Shortcut (ENUMERATION Family)
|
||||
|
||||
You are detecting whether a question exploits option length patterns as a shortcut.
|
||||
|
||||
## Question
|
||||
|
||||
{question}
|
||||
|
||||
## Options
|
||||
|
||||
{options}
|
||||
|
||||
## Designated Correct Answer
|
||||
|
||||
{answer}
|
||||
|
||||
## Instructions
|
||||
|
||||
Check if the question can be answered by exploiting the length or specificity of options, without actually understanding the video content.
|
||||
|
||||
1. Is the correct answer notably longer or more specific than distractors?
|
||||
2. Are distractor options clearly shorter, vaguer, or less detailed than the correct answer?
|
||||
3. Could a student "game" this question by always picking the longest/most-detailed option?
|
||||
|
||||
If option length is an exploitable shortcut → verdict "fail".
|
||||
If options are roughly balanced in length and specificity → verdict "pass".
|
||||
|
||||
## Response Format (strict JSON)
|
||||
|
||||
```json
|
||||
{{"verdict": "pass" or "fail", "reason": "brief explanation"}}
|
||||
```
|
||||
@@ -0,0 +1,32 @@
|
||||
# Leak Test: Frequency Shortcut (REASONING Family)
|
||||
|
||||
You are detecting whether a question exploits word frequency patterns as a shortcut.
|
||||
|
||||
## Question
|
||||
|
||||
{question}
|
||||
|
||||
## Options
|
||||
|
||||
{options}
|
||||
|
||||
## Designated Correct Answer
|
||||
|
||||
{answer}
|
||||
|
||||
## Instructions
|
||||
|
||||
Check if the question can be answered by picking the option that shares the most words or phrases with the question stem, without genuine causal/logical reasoning.
|
||||
|
||||
1. Does the correct answer have significantly more word overlap with the question than distractors?
|
||||
2. Are distractor options phrased in notably different vocabulary from the question?
|
||||
3. Could a student "game" this question by matching keywords between question and options?
|
||||
|
||||
If frequency-based word matching is an exploitable shortcut → verdict "fail".
|
||||
If the question requires genuine reasoning → verdict "pass".
|
||||
|
||||
## Response Format (strict JSON)
|
||||
|
||||
```json
|
||||
{{"verdict": "pass" or "fail", "reason": "brief explanation"}}
|
||||
```
|
||||
@@ -0,0 +1,32 @@
|
||||
# Leak Test: Temporal Proximity Shortcut (RETRIEVAL Family)
|
||||
|
||||
You are detecting whether a question exploits temporal proximity patterns as a shortcut.
|
||||
|
||||
## Question
|
||||
|
||||
{question}
|
||||
|
||||
## Options
|
||||
|
||||
{options}
|
||||
|
||||
## Designated Correct Answer
|
||||
|
||||
{answer}
|
||||
|
||||
## Instructions
|
||||
|
||||
Check if the question can be answered by simply picking the event that is temporally closest to the question's time reference, without truly understanding the content.
|
||||
|
||||
1. Does the correct answer correspond to the most recently mentioned or chronologically nearest event?
|
||||
2. Are distractor options clearly from distant time points, making elimination trivial?
|
||||
3. Could a student "game" this question by always picking the temporally proximate option?
|
||||
|
||||
If temporal proximity is an exploitable shortcut → verdict "fail".
|
||||
If the question requires genuine content understanding → verdict "pass".
|
||||
|
||||
## Response Format (strict JSON)
|
||||
|
||||
```json
|
||||
{{"verdict": "pass" or "fail", "reason": "brief explanation"}}
|
||||
```
|
||||
@@ -0,0 +1,32 @@
|
||||
# Leak Test: Spatial Default Shortcut (SPATIAL Family)
|
||||
|
||||
You are detecting whether a question exploits spatial default assumptions as a shortcut.
|
||||
|
||||
## Question
|
||||
|
||||
{question}
|
||||
|
||||
## Options
|
||||
|
||||
{options}
|
||||
|
||||
## Designated Correct Answer
|
||||
|
||||
{answer}
|
||||
|
||||
## Instructions
|
||||
|
||||
Check if the question can be answered by relying on common spatial assumptions (e.g., "center of frame", "left to right", "foreground") without actual visual understanding.
|
||||
|
||||
1. Is the correct answer the "default" spatial position people would assume (e.g., center, front)?
|
||||
2. Are distractor options in positions that seem intuitively unlikely (e.g., extreme edges, behind)?
|
||||
3. Could a student "game" this question by always picking the spatially default/expected option?
|
||||
|
||||
If spatial defaults are an exploitable shortcut → verdict "fail".
|
||||
If the question requires genuine spatial reasoning → verdict "pass".
|
||||
|
||||
## Response Format (strict JSON)
|
||||
|
||||
```json
|
||||
{{"verdict": "pass" or "fail", "reason": "brief explanation"}}
|
||||
```
|
||||
@@ -0,0 +1,32 @@
|
||||
# Leak Test: Visual Salience Shortcut (VISUAL Family)
|
||||
|
||||
You are detecting whether a question exploits visual salience patterns as a shortcut.
|
||||
|
||||
## Question
|
||||
|
||||
{question}
|
||||
|
||||
## Options
|
||||
|
||||
{options}
|
||||
|
||||
## Designated Correct Answer
|
||||
|
||||
{answer}
|
||||
|
||||
## Instructions
|
||||
|
||||
Check if the question can be answered by simply picking the most visually salient or "obvious" object/action, without careful visual analysis.
|
||||
|
||||
1. Is the correct answer the most prominent or commonly expected element in such a scene?
|
||||
2. Are distractor options obviously implausible or uncommon objects/actions for the described setting?
|
||||
3. Could a student "game" this question by guessing the most visually dominant element?
|
||||
|
||||
If visual salience is an exploitable shortcut → verdict "fail".
|
||||
If the question requires careful visual discrimination → verdict "pass".
|
||||
|
||||
## Response Format (strict JSON)
|
||||
|
||||
```json
|
||||
{{"verdict": "pass" or "fail", "reason": "brief explanation"}}
|
||||
```
|
||||
@@ -0,0 +1,28 @@
|
||||
# Multi-True Gate
|
||||
|
||||
You are checking whether a multiple-choice question has **more than one plausibly correct answer** given the source material.
|
||||
|
||||
## Source Material
|
||||
|
||||
{source_text}
|
||||
|
||||
## Question
|
||||
|
||||
{question}
|
||||
|
||||
## Options
|
||||
|
||||
{options}
|
||||
|
||||
## Instructions
|
||||
|
||||
1. Read the source material and the question carefully.
|
||||
2. For each option, assess whether it could be considered a correct or plausible answer given the source.
|
||||
3. If exactly ONE option is clearly correct → verdict "pass".
|
||||
4. If TWO or more options are plausibly correct → verdict "fail".
|
||||
|
||||
## Response Format (strict JSON)
|
||||
|
||||
```json
|
||||
{{"verdict": "pass" or "fail", "reason": "brief explanation listing plausible options if multiple"}}
|
||||
```
|
||||
@@ -0,0 +1,312 @@
|
||||
"""轻量四门质量检查单元测试。"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from typing import Any
|
||||
|
||||
import pytest
|
||||
|
||||
from app.question_gen.families import RETRIEVAL_FAMILY
|
||||
from app.question_gen.gates import (
|
||||
CandidateQuestion,
|
||||
GateVerdict,
|
||||
_gate_blind_answer,
|
||||
_gate_key_verify,
|
||||
_gate_leak_test,
|
||||
_gate_multi_true,
|
||||
run_gates,
|
||||
)
|
||||
from app.question_gen.postprocess import PostprocessResult
|
||||
from app.tree.index import (
|
||||
IndexMeta,
|
||||
L1Card,
|
||||
L1Node,
|
||||
L2Card,
|
||||
L2Node,
|
||||
L3Card,
|
||||
L3Node,
|
||||
TreeIndex,
|
||||
)
|
||||
from core.types import LLMResponse
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Fixtures
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _make_llm_response(verdict: str, reason: str) -> LLMResponse:
|
||||
"""构造一个 LLM 返回值，content 为标准 JSON 格式。"""
|
||||
return LLMResponse(
|
||||
content=json.dumps({"verdict": verdict, "reason": reason}),
|
||||
thinking="",
|
||||
model="mock-model",
|
||||
provider="mock",
|
||||
prompt_tokens=10,
|
||||
completion_tokens=5,
|
||||
latency_ms=100,
|
||||
ttft_ms=None,
|
||||
max_inter_token_ms=None,
|
||||
cache_hit=False,
|
||||
call_id="mock-call-001",
|
||||
)
|
||||
|
||||
|
||||
class MockLLM:
|
||||
"""可配置的 LLM mock — 按调用顺序返回预设响应。"""
|
||||
|
||||
def __init__(self, responses: list[LLMResponse]) -> None:
|
||||
self._responses = list(responses)
|
||||
self._call_count = 0
|
||||
self.calls: list[dict[str, Any]] = []
|
||||
|
||||
async def chat(
|
||||
self,
|
||||
messages: list[dict[str, Any]],
|
||||
*,
|
||||
session_id: str | None = None,
|
||||
parent_call_id: str | None = None,
|
||||
) -> LLMResponse:
|
||||
"""记录调用并返回预设响应。"""
|
||||
self.calls.append({"messages": messages, "session_id": session_id})
|
||||
idx = self._call_count
|
||||
self._call_count += 1
|
||||
if idx < len(self._responses):
|
||||
return self._responses[idx]
|
||||
# 默认返回 PASS
|
||||
return _make_llm_response("pass", "default")
|
||||
|
||||
|
||||
@pytest.fixture()
|
||||
def mock_llm_pass() -> MockLLM:
|
||||
"""返回始终 PASS 的 mock LLM。"""
|
||||
return MockLLM([_make_llm_response("pass", "evidence found")])
|
||||
|
||||
|
||||
@pytest.fixture()
|
||||
def mock_llm_fail() -> MockLLM:
|
||||
"""返回始终 FAIL 的 mock LLM。"""
|
||||
return MockLLM([_make_llm_response("fail", "no evidence")])
|
||||
|
||||
|
||||
def _make_candidate() -> CandidateQuestion:
|
||||
"""构造测试用候选题目。"""
|
||||
return CandidateQuestion(
|
||||
question_id="q-001",
|
||||
video_id="v-001",
|
||||
task_type="Action Reasoning",
|
||||
skill_target="M1",
|
||||
question="What does the person do after picking up the book?",
|
||||
options=(
|
||||
"A. Reads it",
|
||||
"B. Puts it back",
|
||||
"C. Throws it away",
|
||||
"D. Gives it to someone",
|
||||
),
|
||||
answer="A",
|
||||
source_nodes=("L2_001", "L3_001"),
|
||||
difficulty="medium",
|
||||
)
|
||||
|
||||
|
||||
def _make_tree() -> TreeIndex:
|
||||
"""构造最小测试树。"""
|
||||
l3_card = L3Card(
|
||||
frame_summary="Person picks up a book from the shelf and starts reading it.",
|
||||
visible_entities=["person", "book", "shelf"],
|
||||
ongoing_actions=["picking up book", "reading"],
|
||||
visible_text=[],
|
||||
spatial_layout="person in center, shelf on left",
|
||||
visual_attributes={},
|
||||
subtitle="He picks up the book and reads.",
|
||||
)
|
||||
l3 = L3Node(id="L3_001", card=l3_card)
|
||||
|
||||
l2_card = L2Card(
|
||||
event_description="A person picks up a book from the shelf and begins reading it attentively.",
|
||||
entities=["person", "book"],
|
||||
actions=["pick up", "read"],
|
||||
action_subjects=["person"],
|
||||
visible_text=[],
|
||||
spatial_relations="person near shelf",
|
||||
state_changes="book moves from shelf to hands",
|
||||
subtitle="He picks up the book and reads.",
|
||||
)
|
||||
l2 = L2Node(id="L2_001", card=l2_card, children=[l3])
|
||||
|
||||
l1_card = L1Card(
|
||||
scene_summary="Library scene with a person browsing and reading books.",
|
||||
main_setting="library",
|
||||
key_entities=["person", "books", "shelf"],
|
||||
main_actions=["browsing", "reading"],
|
||||
topic_keywords=["library", "reading"],
|
||||
visible_text=[],
|
||||
temporal_flow="enter → browse → pick up → read",
|
||||
)
|
||||
l1 = L1Node(id="L1_001", card=l1_card, children=[l2])
|
||||
|
||||
meta = IndexMeta(source_path="test.mp4", modality="video")
|
||||
return TreeIndex(metadata=meta, roots=[l1])
|
||||
|
||||
|
||||
def _make_postprocess(verbatim_ratio: float = 0.1) -> PostprocessResult:
|
||||
"""构造测试用后处理结果。"""
|
||||
return PostprocessResult(
|
||||
options=(
|
||||
"A. Reads it",
|
||||
"B. Puts it back",
|
||||
"C. Throws it away",
|
||||
"D. Gives it to someone",
|
||||
),
|
||||
answer="A",
|
||||
referent_violations=[],
|
||||
verbatim_ratio=verbatim_ratio,
|
||||
has_time_anchor=False,
|
||||
)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# TestGateKeyVerify
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class TestGateKeyVerify:
|
||||
"""key_verify 门：验证答案在来源素材中有证据支撑。"""
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_pass_evidence(self, mock_llm_pass: MockLLM) -> None:
|
||||
"""LLM 确认有证据 → PASS。"""
|
||||
candidate = _make_candidate()
|
||||
tree = _make_tree()
|
||||
result = await _gate_key_verify(candidate, tree, mock_llm_pass, session_id="test-session")
|
||||
assert result.verdict == GateVerdict.PASS
|
||||
assert mock_llm_pass.calls # 确认调用了 LLM
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_fail_no_evidence(self, mock_llm_fail: MockLLM) -> None:
|
||||
"""LLM 判断无证据 → FAIL。"""
|
||||
candidate = _make_candidate()
|
||||
tree = _make_tree()
|
||||
result = await _gate_key_verify(candidate, tree, mock_llm_fail, session_id="test-session")
|
||||
assert result.verdict == GateVerdict.FAIL
|
||||
assert "no evidence" in result.reason
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# TestGateBlindAnswer
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class TestGateBlindAnswer:
|
||||
"""blind_answer 门：无上下文答对 → FAIL（题目太简单/泄漏）。"""
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_pass_wrong(self) -> None:
|
||||
"""LLM 在无上下文时答错 → PASS（题目确实需要视频信息）。"""
|
||||
# LLM 返回 "pass" 表示它无法正确回答
|
||||
llm = MockLLM([_make_llm_response("pass", "cannot determine without context")])
|
||||
candidate = _make_candidate()
|
||||
result = await _gate_blind_answer(candidate, llm, session_id="test-session")
|
||||
assert result.verdict == GateVerdict.PASS
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_fail_correct(self) -> None:
|
||||
"""LLM 在无上下文时答对 → FAIL（题目泄漏）。"""
|
||||
llm = MockLLM([_make_llm_response("fail", "answer is obvious from options")])
|
||||
candidate = _make_candidate()
|
||||
result = await _gate_blind_answer(candidate, llm, session_id="test-session")
|
||||
assert result.verdict == GateVerdict.FAIL
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# TestGateMultiTrue
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class TestGateMultiTrue:
|
||||
"""multi_true 门：多选项正确 → FAIL。"""
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_pass_single(self) -> None:
|
||||
"""只有一个正确选项 → PASS。"""
|
||||
llm = MockLLM([_make_llm_response("pass", "only one correct answer")])
|
||||
candidate = _make_candidate()
|
||||
tree = _make_tree()
|
||||
result = await _gate_multi_true(candidate, tree, llm, session_id="test-session")
|
||||
assert result.verdict == GateVerdict.PASS
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_fail_multi(self) -> None:
|
||||
"""多个选项可被视为正确 → FAIL。"""
|
||||
llm = MockLLM([_make_llm_response("fail", "options A and B are both plausible")])
|
||||
candidate = _make_candidate()
|
||||
tree = _make_tree()
|
||||
result = await _gate_multi_true(candidate, tree, llm, session_id="test-session")
|
||||
assert result.verdict == GateVerdict.FAIL
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# TestGateLeakTest
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class TestGateLeakTest:
|
||||
"""leak_test 门：按家族模板执行泄漏探测。"""
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_per_family_template(self) -> None:
|
||||
"""使用家族特定的 probe_template 调用 LLM。"""
|
||||
llm = MockLLM([_make_llm_response("pass", "no shortcut detected")])
|
||||
candidate = _make_candidate()
|
||||
result = await _gate_leak_test(candidate, RETRIEVAL_FAMILY, llm, session_id="test-session")
|
||||
assert result.verdict == GateVerdict.PASS
|
||||
# 验证 session_id 被正确传递
|
||||
assert llm.calls[0]["session_id"] == "test-session"
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# TestRunGates
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class TestRunGates:
|
||||
"""run_gates 编排：并发四门 + verbatim 短路。"""
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_all_pass(self) -> None:
|
||||
"""四门全 PASS → GateReport.passed=True。"""
|
||||
llm = MockLLM([_make_llm_response("pass", f"gate {i} ok") for i in range(4)])
|
||||
candidate = _make_candidate()
|
||||
tree = _make_tree()
|
||||
postprocess = _make_postprocess(verbatim_ratio=0.1)
|
||||
|
||||
report = await run_gates(
|
||||
candidate=candidate,
|
||||
tree=tree,
|
||||
llm=llm,
|
||||
family_spec=RETRIEVAL_FAMILY,
|
||||
postprocess=postprocess,
|
||||
session_id="test-session",
|
||||
)
|
||||
assert report.passed is True
|
||||
assert report.reject_reason is None
|
||||
|
||||
@pytest.mark.asyncio()
|
||||
async def test_high_verbatim_shortcircuits(self) -> None:
|
||||
"""verbatim_ratio > 0.5 → key_verify 直接 FAIL，不调用 LLM。"""
|
||||
llm = MockLLM([_make_llm_response("pass", "should not be called")])
|
||||
candidate = _make_candidate()
|
||||
tree = _make_tree()
|
||||
postprocess = _make_postprocess(verbatim_ratio=0.8)
|
||||
|
||||
report = await run_gates(
|
||||
candidate=candidate,
|
||||
tree=tree,
|
||||
llm=llm,
|
||||
family_spec=RETRIEVAL_FAMILY,
|
||||
postprocess=postprocess,
|
||||
session_id="test-session",
|
||||
)
|
||||
assert report.passed is False
|
||||
assert report.key_verify.verdict == GateVerdict.FAIL
|
||||
assert "verbatim" in report.key_verify.reason.lower()
|
||||
Reference in New Issue
Block a user