Dataset Card for tw-iso-27001-bench
⚠️ 狀態:manual review 中 —— 本資料集為 LLM 自動生成,尚未經過人工逐題審核,使用前請自行抽查驗證。
📋 tw-iso-27001-bench 是以 lianghsun/tw-cybersecurity 為來源,自動生成的 100 題繁體中文資安單選題評測集,每題皆附上足以回答該題的上下文(context),可直接用於評測語言模型的資安知識理解能力。
Dataset Details
Dataset Description
本資料集從 lianghsun/tw-cybersecurity 的三個子集(iso-27001、practices、threat-intel-and-defense)分層抽樣 100 篇文件,並用 LLM 針對每篇文件生成一題單選題。每題的 context 欄位是從原文擷取「足以回答該題」的段落,讓評測時不需要額外提供外部知識,只憑 context 就能作答。
- Curated by: Liang Hsun Huang
- Language(s) (NLP): Traditional Chinese
- License: MIT
Dataset Sources
- Repository: lianghsun/tw-iso-27001-bench
- Source dataset: lianghsun/tw-cybersecurity
Uses
Direct Use
- 評測語言模型對繁體中文資安(ISO 27001、資安實務、威脅情資)知識的理解與判讀能力。
- 作為 few-shot 或 in-context QA 評測的單選題來源。
Out-of-Scope Use
- 不建議直接作為訓練資料使用(未經人工審核,可能有誤)。
- 不適用於需要嚴格正確性保證的正式資安認證測驗。
Dataset Structure
id, category, source_file, context, question, A, B, C, D, answer
id:題號(0–99)。category:來源子集,iso-27001/practices/threat-intel-and-defense。source_file:來源文件的檔名或識別碼。context:從來源文件擷取、足以回答該題的段落。question/A/B/C/D:題目與四個選項。answer:正確答案(A/B/C/D)。
三個子集抽樣數量:iso-27001 17 題(全部)、practices 33 題、threat-intel-and-defense 50 題。
Dataset Creation
Curation Rationale
現有的 lianghsun/tw-cybersecurity 是長文件與問答格式,缺乏標準化的單選題評測集。本資料集將其轉換為附完整上下文的單選題,方便直接用於模型評測。
Source Data
Data Collection and Processing
- 從 lianghsun/tw-cybersecurity 三個子集分層抽樣 100 篇文件(iso-27001 全部 17 篇、practices 33 篇、threat-intel-and-defense 50 篇)。
- 對每篇文件,以 LLM 生成一題單選題(要求輸出 JSON,包含足以回答該題的 context、題目、四個選項、正確答案)。
- 選項順序經隨機重排,避免正確答案集中在特定字母(原始生成結果有明顯偏向 A/B 的現象,已重排改善)。
Who are the source data producers?
原始文件來自 lianghsun/tw-cybersecurity;題目由 LLM 自動生成,未經人工撰寫。
Annotations
本資料集的題目、選項、答案皆由 LLM 自動生成,未經人工標註或審核。
Personal and Sensitive Information
本資料集內容為公開資安教育文件與威脅情資摘要,不含個人隱私資訊。
Bias, Risks, and Limitations
- 題目、選項、答案皆由 LLM 自動生成,尚未經過人工審核,可能存在事實錯誤、選項有歧義或標註答案有誤的情形。
- 干擾選項的品質取決於生成模型的判斷,可能存在部分選項過於明顯錯誤(鑑別度不足)的情況。
Recommendations
使用前建議自行抽查驗證題目與答案的正確性,尤其若用於正式評測或發布結果。
Citation
@misc{tw_iso_27001_bench,
title = {tw-iso-27001-bench: A Traditional Chinese Cybersecurity Single-Choice Evaluation Set},
author = {Huang, Liang Hsun},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-iso-27001-bench}},
note = {Generated from lianghsun/tw-cybersecurity via LLM.}
}
Dataset Card Authors
Dataset Card Contact
- Downloads last month
- 12