8d515ff01f
算法 100% 保真 TRM4: task_types 过滤、correctness.get(id, False) 语义、 对题在前返回顺序、min_per_class 遍历 pool 全部题型(含稀疏类)。 所有参数显式传入,无默认值。 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
168 lines
5.5 KiB
Python
168 lines
5.5 KiB
Python
"""题目加载与分层采样。
|
|
|
|
从 benchmark JSON 目录加载题目,提供按对错比例的分层采样。
|
|
对应训练循环中的 DataLoader 角色。
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import random
|
|
from typing import TYPE_CHECKING
|
|
|
|
from core.types import GeneratedQuestion
|
|
|
|
if TYPE_CHECKING:
|
|
from pathlib import Path
|
|
|
|
_LEGACY_DEFAULT_DIFFICULTY = "medium"
|
|
|
|
|
|
def load_benchmark(questions_dir: Path) -> list[GeneratedQuestion]:
|
|
"""从 benchmark JSON 目录加载题目列表。
|
|
|
|
每个 JSON 文件以文件名(不含扩展名)作为 video_id,
|
|
文件内容为题目数组。
|
|
|
|
参数:
|
|
questions_dir: 包含 *.json 文件的目录路径。
|
|
|
|
返回:
|
|
按文件名排序加载的题目列表。
|
|
"""
|
|
results: list[GeneratedQuestion] = []
|
|
for path in sorted(questions_dir.glob("*.json")):
|
|
video_id = path.stem
|
|
with open(path, encoding="utf-8") as f:
|
|
qa_list: list[dict] = json.load(f)
|
|
for qa in qa_list:
|
|
results.append(
|
|
GeneratedQuestion(
|
|
question_id=qa["question_id"],
|
|
video_id=video_id,
|
|
task_type=qa["task_type"],
|
|
question=qa["question"],
|
|
options=tuple(qa["options"]),
|
|
answer=qa["answer"],
|
|
source_nodes=tuple(qa.get("source_nodes", ())),
|
|
difficulty=qa.get("difficulty", _LEGACY_DEFAULT_DIFFICULTY),
|
|
)
|
|
)
|
|
return results
|
|
|
|
|
|
def stratified_sample(
|
|
questions: list[GeneratedQuestion],
|
|
correctness: dict[str, bool],
|
|
size: int,
|
|
correct_ratio: float | None,
|
|
task_types: list[str] | None,
|
|
seed: int,
|
|
min_per_class: int | None,
|
|
) -> list[GeneratedQuestion]:
|
|
"""按题型过滤后采样 size 道题,可选按对错比例分层并按题型保底。
|
|
|
|
参数:
|
|
questions: 候选题目全集。
|
|
correctness: question_id -> 基线是否答对。
|
|
size: 采样总量。
|
|
correct_ratio: 采样中"基线答对"题的占比;None 表示自然分布。
|
|
task_types: 限定题型;None 表示不限。
|
|
seed: 随机种子,保证可复现。
|
|
min_per_class: 每个题型补足到的下限;None 表示不补足。
|
|
|
|
返回:
|
|
采样后的题目列表。
|
|
|
|
异常:
|
|
ValueError: 自然分布时池不足 size,或分层时某层题目不足。
|
|
"""
|
|
rng = random.Random(seed)
|
|
pool = [q for q in questions if task_types is None or q.task_type in task_types]
|
|
|
|
if correct_ratio is None:
|
|
if len(pool) < size:
|
|
raise ValueError(f"自然分布采样不足: 需 {size} 道, 实有 {len(pool)} 道")
|
|
sampled = rng.sample(pool, size)
|
|
else:
|
|
sampled = _ratio_stratified_sample(pool, correctness, size, correct_ratio, rng)
|
|
|
|
if min_per_class is not None:
|
|
sampled = _backfill_per_class(sampled, pool, min_per_class, rng)
|
|
return sampled
|
|
|
|
|
|
def _ratio_stratified_sample(
|
|
pool: list[GeneratedQuestion],
|
|
correctness: dict[str, bool],
|
|
size: int,
|
|
correct_ratio: float,
|
|
rng: random.Random,
|
|
) -> list[GeneratedQuestion]:
|
|
"""按对错比例分层采样:对题占 correct_ratio,其余为错题。
|
|
|
|
参数:
|
|
pool: 题型过滤后的候选题。
|
|
correctness: question_id -> 基线是否答对。
|
|
size: 采样总量。
|
|
correct_ratio: 对题占比。
|
|
rng: 随机数发生器。
|
|
|
|
返回:
|
|
采样后的题目列表(对题在前、错题在后)。
|
|
|
|
异常:
|
|
ValueError: 对题或错题层不足。
|
|
"""
|
|
correct = [q for q in pool if correctness.get(q.question_id, False)]
|
|
wrong = [q for q in pool if not correctness.get(q.question_id, False)]
|
|
n_correct = round(size * correct_ratio)
|
|
n_wrong = size - n_correct
|
|
if len(correct) < n_correct or len(wrong) < n_wrong:
|
|
raise ValueError(
|
|
f"分层不足: 需对{n_correct}/错{n_wrong}, 实有对{len(correct)}/错{len(wrong)}"
|
|
)
|
|
return rng.sample(correct, n_correct) + rng.sample(wrong, n_wrong)
|
|
|
|
|
|
def _backfill_per_class(
|
|
sampled: list[GeneratedQuestion],
|
|
pool: list[GeneratedQuestion],
|
|
min_per_class: int,
|
|
rng: random.Random,
|
|
) -> list[GeneratedQuestion]:
|
|
"""对候选池中出现的每个题型,将采样结果补足到 min_per_class 道。
|
|
|
|
遍历对象是候选池 pool 里出现的全部题型(非仅 sampled 命中的),
|
|
保证任意稀疏题型都能拿到足额样本。
|
|
|
|
参数:
|
|
sampled: 主采样结果(不修改,返回新列表)。
|
|
pool: 候选题全集(补足来源 + 题型枚举来源)。
|
|
min_per_class: 每个题型的下限。
|
|
rng: 随机数发生器。
|
|
|
|
返回:
|
|
补足后的题目列表。
|
|
"""
|
|
selected_ids = {q.question_id for q in sampled}
|
|
result = list(sampled)
|
|
counts: dict[str, int] = {}
|
|
for q in sampled:
|
|
counts[q.task_type] = counts.get(q.task_type, 0) + 1
|
|
ordered_task_types: dict[str, None] = {}
|
|
for q in pool:
|
|
ordered_task_types.setdefault(q.task_type, None)
|
|
for task_type in ordered_task_types:
|
|
deficit = min_per_class - counts.get(task_type, 0)
|
|
if deficit <= 0:
|
|
continue
|
|
candidates = [
|
|
q for q in pool if q.task_type == task_type and q.question_id not in selected_ids
|
|
]
|
|
take = rng.sample(candidates, min(deficit, len(candidates)))
|
|
for q in take:
|
|
selected_ids.add(q.question_id)
|
|
result.append(q)
|
|
return result
|