Mana29/SakuraQA_dataset

SakuraQA: A Japanese Youth Culture Benchmark for LLMs

SakuraQA is a 4-choice question answering benchmark specifically designed to evaluate the cultural adaptability and understanding of Large Language Models (LLMs) regarding Japanese youth culture and contemporary trends.

The dataset was curated by 132 university students from Japan Women's University, reflecting authentic, first-person perspectives, slangs, and trends up to June 2025.

---

Dataset Statistics

SakuraQA consists of 410 multiple-choice questions across 5 distinct categories:

Category	Japanese Name	Questions	Description
Subculture	サブカルチャー	117	Anime, manga, games, popular characters, storylines, and game mechanics.
Fashion & Beauty	おしゃれ	47	Fashion trends, apparel brands, cosmetics, makeup, and hairstyling.
Gourmet	食べ物	62	Trendy foods, culinary knowledge, ingredients, and popular dining spots.
Slang & Buzzwords	流行語	64	Youth slang, internet memes, and social media expressions.
Entertainment	エンタメ	120	Music, TV shows, movies, live events, idols, and theme parks.
Total		410

---

Data Structure

Each sample is formatted as a single JSON object per line (JSONL):

{
  "question_id": 1,
  "category": "subculture",
  "category_ja": "サブカルチャー",
  "question": "『わんだふるぷりきゅあ!』で3番目に変身したプリキュアは?",
  "choices": {
    "A": "キュアフレンディ",
    "B": "キュアリリン",
    "C": "キュアワンダフル",
    "D": "キュアニャミー"
  },
  "answer": "D"
}

Fields

• question_id (int): Unique identifier for the question.
• category (str): English category name (subculture, fashion, gourmet, slang, entertainment).
• category_ja (str): Japanese category name.
• question (str): The question text.
• choices (dict): 4 candidate options (A, B, C, D).
• answer (str): Ground-truth answer label (A, B, C, or D).

---

Benchmark Results (JSAI 2026)

All models were evaluated under 0-shot ($temperature=0$) conditions:

Closed Models

• Gemini (gemini-2.5-flash): Consistently high performance across all categories (Maintained >70% accuracy even in the slang category).
• Claude (claude-sonnet-4-20250514): Achieved the highest score in the Gourmet category (92%).
• GPT-4o (gpt-4o-mini): Strong in Food and Fashion, but struggled with niche youth slang and fine-grained entertainment knowledge.

Open-source Models

• Swallow (Llama-3.1-Swallow-8B-Instruct-v0.5): Exhibited balanced performance across diverse Japanese cultural contexts.
• Sarashina (sarashina2.2-3b-instruct-v0.1): High accuracy in Food and Fashion domains.
• CALM (calm2-7b-chat): Showed distinct strengths in the Entertainment category compared to other open models.

---

Quick Start

Python Standard

import json

def load_sakura_qa(filepath="data/sakura_qa_all.jsonl"):
    with open(filepath, "r", encoding="utf-8") as f:
        return [json.loads(line) for line in f]

dataset = load_sakura_qa()
print(f"Loaded {len(dataset)} questions.")
print(dataset[0])

Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("json", data_files="data/sakura_qa_all.jsonl")
print(dataset["train"][0])

---

Citation

If you use SakuraQA in your research, please cite the following paper:

@inproceedings{hasegawa2026sakuraqa,
  title     = {若者にとって本当に使えるAIとは何か: 若者文化ベンチマーク SakuraQA の構築},
  author    = {長谷川 愛珠 and 宮田 侑佳 and 伊東 和香 and 小田垣 真加 and 西潟 優羽 and 倉光 君郎},
  booktitle = {2026年度 人工知能学会全国大会 (第40回) 論文集},
  pages     = {5Yin-A-57},
  year      = {2026},
  publisher = {人工知能学会}
}

License

This dataset is distributed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.