LianeMarilin/long-context-qa-curated-20

Dataset Card / 数据集卡

Dataset Description / 数据集简介

This public release contains 20 curated samples selected from a 10,000-record long-context QA collection. It targets retrieval over long documents, cross-section evidence synthesis, numerical reasoning, timeline reconstruction, and structured answer evaluation. The public subset contains 15 short-answer questions and 5 multiple-choice questions, balanced across Chinese and English.

本公开版本从 10,000 条长上下文问答数据中精选 20 条，用于长文档检索、跨段证据整合、数值推理、时序重建和结构化回答评测；其中简答题 15 道、选择题 5 道，中英文各 10 道。

• Task: long-context question answering and LLM evaluation
• Languages: Chinese (zh) and English (en)
• Public size: 20 QA rows, 160 associated rollout records
• Full collection: 10,000 records
• Context range: approximately 8k–256k length buckets
• Source repository: GitHub

Dataset Structure and Splits / 数据结构与划分

Config	Viewer split	Rows	File
default	train	20	data/long_context_qa_curated_20.jsonl

train is used as the Viewer container split. The 20-row public release is an evaluation sample rather than a prescribed model-training partition. Context documents, rollout traces, scores, and validation artifacts remain as individual files in the repository.

Data Fields / 字段说明

Field	Description
question_id, context_id	Stable question and context identifiers
language, domain, document_type	Language and content taxonomy
context, question, answer	Long source context, prompt, and reference answer
question_type, choices	Short-answer or multiple-choice type and options
evidence, answer_explanation	Supporting evidence and answer rationale
context_length_tier, length_bucket	Context-length grouping
rollouts, difficulty_result	Independent model attempts and aggregated grading
source_url, file_path, rollout_path	Provenance and repository paths

Loading / 加载方式

from datasets import load_dataset

ds = load_dataset("LianeMarilin/long-context-qa-curated-20")
print(ds["train"][0])

The JSONL is also directly readable with any line-oriented JSON parser; no archive extraction step is required.

Curation, Intended Uses, and Limitations / 筛选、用途与局限

Samples were selected for bilingual balance, answerability, evidence traceability, question-type coverage, and difficulty. Suitable uses include long-context QA evaluation, retrieval analysis, judge calibration, and error analysis. Report results separately by language, question type, domain, and context-length bucket. The public subset is small and intentionally difficult, so it should not be treated as a population-level estimate or as a complete training corpus. Source-page availability and licensing may vary; users should review each record's provenance fields before redistribution.

License and Citation / 许可与引用

Dataset metadata uses license: other: benchmark-authored annotations and scripts may be reused with attribution, while linked source material remains governed by its original terms. Cite the dataset repository and preserve record-level provenance.

@misc{long_context_qa_curated_20_2026,
  title  = {Long Context QA: Curated 20},
  author = {Liane Marilin},
  year   = {2026},
  url    = {https://huggingface.co/datasets/LianeMarilin/long-context-qa-curated-20}
}

Long Context QA · Curated 20

面向长文档检索、跨段证据整合与结构化推理的精选问答数据集
A compact, inspectable benchmark for long-context question answering.

---

✨ 数据集亮点

	特性	说明
🧠	真实长上下文	8k—256k长度桶，覆盖局部检索、跨章节聚合、时序重建与数值推理
🎯	精选20题	15道简答题 + 5道选择题，纳入源数据中全部有效选择题
🌏	双语均衡	中文10题、英文10题
🧪	可复核难度	每题保留8次独立rollout、原始回答、评分与裁判过程
🔗	证据可追溯	每条记录关联独立context文件、来源链接与答案证据
✅	机器可验证	提供零依赖校验脚本，检查JSONL、文件映射、ID、题型与rollout

📊 数据概览

指标	数值
精选题目	20
简答 / 选择	15 / 5
中文 / 英文	10 / 10
独立rollout	160
整体平均正确率	14.37%
长度桶	8k · 16k · 32k · 64k · 128k · 256k
覆盖领域	软件工程、金融、学术、法律、政府事务、新闻

查看领域分布

领域	题数
软件与工程	6
金融	5
法律	3
政府事务	3
学术	2
新闻	1

🧭 评测链路

flowchart LR
    A[Long Context] --> B[Question]
    B --> C[8 Independent Rollouts]
    C --> D[Semantic Judge]
    D --> E[Per-run Scores]
    E --> F[Average Accuracy]
    A --> G[Evidence]
    G --> D

🗂️ 仓库结构

.
├── assets/                         # README视觉素材
├── configs/                        # 三批rollout配置
├── contexts/                       # 20份长上下文原文
├── data/
│   ├── long_context_qa_curated_20.jsonl
│   ├── dataset_card.json
│   └── index.csv
├── docs/
│   └── selection-report.md         # 逐题精选清单
├── rollouts/                       # 每题8轮回答、评分与裁判
├── scripts/
│   └── validate_dataset.py
└── README.md

🚀 快速开始

读取数据

import json
from pathlib import Path

path = Path("data/long_context_qa_curated_20.jsonl")
records = [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines()]

print(f"records: {{len(records)}}")
print(records[0]["question"])
print(records[0]["answer"])

运行完整校验

python3 scripts/validate_dataset.py

成功时输出：

PASS · 20 records · 20 contexts · 160 rollouts

🧩 核心字段

字段	类型	含义
track_id	string	稳定题目ID
context_id	string	长上下文ID
context	string	完整上下文文本
question_type	string	短答案题、选择题或多项选择题
question	string	问题及输出约束
answer	array	标准答案
answer_explanation	object	解题步骤与证据
review_result	object	内容质量审查
difficulty_result	object	8轮回答、逐轮评分与平均结果
task_labels	object	主任务、次任务、上下文需求等级

🏅 精选策略

• 排除源包中已明确标记为证据不足或答案不唯一的题目。
• 相同context只保留一道题，降低重复度。
• 纳入源数据中全部5道有效选择题。
• 在剩余题目中平衡语言、领域、长度桶与任务类型。
• 保留原track_id与context_id，便于回溯。
• 修正000026的JSON标准答案和000037的题型元数据。

完整记录见 精选题清单。

🔍 单条记录示例

{{
  "track_id": "longqa_delivery_000001",
  "context_id": "context_000001",
  "question_type": "短答案题",
  "language": "Chinese",
  "domain": "软件与工程",
  "token_length": 8488,
  "file_path": "contexts/context_000001.txt",
  "question": "...",
  "answer": ["..."],
  "review_result": {{"status": "pass"}},
  "difficulty_result": {{"rollout_count": 8}}
}}

🛠️ 已修订记录

ID	修订内容
000026	标准答案改成题干指定的[Answer] + JSON数组结构
000037	question_type由“短答案题”修正为“选择题”

📌 使用说明

完整数据集共 10,000条，本仓库从中精选 20条高质量样例，用于展示数据格式、任务类型、长上下文推理难度及完整评测流程。

---

Built for inspectable long-context evaluation.
Context → Evidence → Reasoning → Answer