llm-jp/MECHA-ja-v2

MECHA-Ja v2: Multimodal Everyday-life and Cultural Habits Assessment for Japanese

🤗 MECHA-Ja (v1)

Overview

MECHA-Ja v2 is a multiple-choice (four-option) VQA benchmark on Japanese everyday life and cultural habits. It has 3,994 questions based on 1,519 images; rather than the names of depicted objects, it asks what they mean in Japanese life and culture and how they are handled. Each question is labeled as either a Cultural Common Knowledge question, answerable with only knowledge naturally acquired through life and basic education in Japan, or a Cultural Encyclopedic Knowledge question, requiring additional learning about specific subjects beyond that. This enables separate evaluation of the two types of knowledge.

MECHA-Ja v2 builds on MECHA-Ja (v1) and extends it with additional questions.

What's New in v2

We extended v1 with new questions, bringing the total to 3,994. Every question is now labeled with its knowledge type, enabling separate evaluation of Cultural Common Knowledge and Cultural Encyclopedic Knowledge. We also added 70 multiple-answer questions, in which all correct options must be selected. For differences in the fields, see Fields.

Construction Pipeline

MECHA-Ja v2 extends v1 with newly created questions through the following four stages:

1. Image Collection: We newly collected 917 images taken in Japan, showing everyday objects, scenery, and cultural customs. Faces are blurred in images that could identify a person.
2. Question Creation: Annotators wrote a question, four answer options, and the correct answer(s) about Japanese everyday life and cultural habits, adding 2,186 questions.
3. Integration with v1: The new questions were integrated with the v1 questions. Questions that did not meet our design guidelines, or whose correct answers could not be confirmed, were excluded.
4. Labeling: All 3,994 questions are labeled as Cultural Common Knowledge or Cultural Encyclopedic Knowledge.

Usage

from datasets import load_dataset

ds = load_dataset("llm-jp/MECHA-ja-v2", split="test")
sample = ds[0]
image = sample["image"]
question = sample["question"]
options = sample["options"]
answer = sample["answer"]  # list of class indices; 0-3 correspond to options 1-4

You can filter by knowledge_type to separate Cultural Common Knowledge and Cultural Encyclopedic Knowledge questions.

label = ds.features["knowledge_type"]

# Cultural Common Knowledge questions
common = ds.filter(lambda x: x["knowledge_type"] == label.str2int("common"))

# Cultural Encyclopedic Knowledge questions
encyclopedic = ds.filter(lambda x: x["knowledge_type"] == label.str2int("encyclopedic"))

Fields

• id: question ID.
• question: question text.
• answer_cardinality: number of correct options. single (one) or multi (two or three).
• knowledge_type: type of knowledge required to answer, classified by how that knowledge is acquired. common (Cultural Common Knowledge): answerable with only knowledge naturally acquired through life and basic education in Japan. encyclopedic (Cultural Encyclopedic Knowledge): requires additional learning about specific subjects beyond that. The classification is independent of question difficulty.
• background_text: supplementary note that can be referred to when answering, providing information that is not apparent from the image alone or is difficult to grasp from it. Present only in some questions.
• options: list of four answer options.
• image: input image.
• answer: list of correct options. Indices 0-3 correspond to options 1-4. Multiple-answer questions contain multiple elements.
• source_version: version in which the question was created. v1 (carried over from v1) or v2 (newly created in v2).
• answer_type: format of the answer options. Factoid: names of things or single words. Non-Factoid: sentences giving reasons or methods. The format is consistent across the options of a question.
• explanation: brief explanation of the question.
• japanese_score: score for how uniquely Japanese the question's knowledge is (0-1).
• daily_score: everydayness score (0-1).

answer was a ClassLabel in v1 and is a list of ClassLabel in v2.

answer_type, explanation, japanese_score, and daily_score are available only for questions whose source_version is v1. For the definitions of japanese_score and daily_score, see Hasegawa et al. (FIT2025, in Japanese).

License

MECHA-Ja v2 is released under the Apache License 2.0. Images are only those licensed for commercial use and public release.