Dataset Card for Shitsukan Dataset

English | 日本語

• Paper: Evaluating Model Alignment with Human Perception: A Study on Shitsukan in LLMs and LVLMs (COLING 2025)
• Evaluation code: cl-tohoku/shitsukan-eval

Dataset Summary

The Shitsukan Dataset is designed to evaluate the alignment of Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) with human perception, focusing on the Japanese concept of shitsukan (質感).
Shitsukan represents the sensory experience when perceiving objects—an inherently vague and highly subjective concept encompassing texture, material properties, and emotional responses.
This dataset includes shitsukan terms recalled by individuals in response to images of specified objects (the images are from MS COCO 2017). Additionally, benchmark tasks are designed to evaluate the shitsukan recognition capabilities of LLMs and LVLMs.

This repository contains the original human-annotated dataset and three task-specific datasets created by processing this dataset:

• Perception Tasks: perception generation and perception selection (with images)
• Commonsense Tasks: commonsense generation and commonsense classification (text only)
• Taxonomic Tasks: yes-no, A-B, and multiple-choice classification (text only)

Note: All data in this repository is intended for evaluation only and is published as the test split. There are no training or validation splits.

If you want to run the evaluation script using this dataset, please check our GitHub repository (shitsukan-eval).

Languages

The dataset includes content in:

• Japanese (ja)
• English (en)

The English data are translations of the Japanese data. The languages available for each task are as follows:

Task	Sub-task	ja	en	Image input
Perception	Generation	✓		✓
Perception	Selection	✓	✓	✓
Commonsense	Generation	✓
Commonsense	Classification	✓	✓
Taxonomic	Yes-No / A-B / Multiple-choice classification	✓	✓

Usage

Loading with 🤗 Datasets

Each data file is registered as its own configuration, and every configuration has only a test split.

from datasets import load_dataset

# Evaluation data of a task (e.g., perception selection with 2 choices, Japanese)
test = load_dataset("tohoku-nlp/Shitsukan", "perception_selection_choice2_ans1_ja", split="test")

# Few-shot examples for the same task
few_shot = load_dataset("tohoku-nlp/Shitsukan", "perception_selection_choice2_ans1_ja_few_shot", split="test")

# Without a configuration name, the default configuration (original_crowdsourced_and_cleaned) is loaded
original = load_dataset("tohoku-nlp/Shitsukan", split="test")

Using the raw files (e.g., with shitsukan-eval)

The JSONL files are stored with Git LFS:

git lfs install
git clone https://huggingface.co/datasets/tohoku-nlp/Shitsukan

shitsukan-eval reads these JSONL files directly; the task configuration files (shitsukan_eval/tasks/{task}/{sub_task}/{task}_{sub_task}_{lang}.yaml) point to the data/... paths of this repository. The perception tasks also require the MS COCO 2017 images (see the Images section below). Note that shitsukan-eval currently implements the taxonomic tasks for Japanese only.

Dataset Structure

The dataset is organized into several directories and files:

data/
│
├── original_data/
│   ├── crowdsourced_and_cleaned_shitsukan_dataset.jsonl
│   ├── data_annotated_through_taskA_shitsukan_term_selection.jsonl
│   ├── data_annotated_through_taskB_shitsukan_term_naturalness_en.jsonl
│   ├── data_annotated_through_taskB_shitsukan_term_naturalness_ja.jsonl
│   └── no_shitsukan_datasets_excluded_during_cleaning.jsonl
│
├── commonsense/
│   ├── classification/
│   │   ├── commonsense_classification_en.jsonl
│   │   ├── commonsense_classification_ja.jsonl
│   │   ├── few_shot_for_commonsense-classification_en.jsonl
│   │   └── few_shot_for_commonsense-classification_ja.jsonl
│   └── generation/
│       ├── commonsense_generation_ja.jsonl
│       └── few_shot_for_commonsense-generation_ja.jsonl
│
├── perception/
│   ├── generation/
│   │   ├── perception_generation_ja.jsonl
│   │   └── few_shot_for_perception-generation_ja.jsonl
│   └── selection/
│       ├── choice2_ans1/
│       │   ├── perception_selection_choice2-ans1_en.jsonl
│       │   ├── perception_selection_choice2-ans1_ja.jsonl
│       │   ├── few_shot_for_perception-selection-choice2-ans1_en.jsonl
│       │   └── few_shot_for_perception-selection-choice2-ans1_ja.jsonl
│       ├── choice3_ans1/
│       │   └── ...
│       ├── choice4_ans1/
│       │   └── ...
│       ├── choice5_ans1/
│       │   └── ...
│       └── random/
│           └── ...
│
└── taxonomic/
    ├── a_b_classification/
    │   ├── taxonomic_a-b_classification_en.jsonl
    │   ├── taxonomic_a-b_classification_ja.jsonl
    │   ├── few_shot_for_taxonomic_a-b_classification_en.jsonl
    │   └── few_shot_for_taxonomic_a-b_classification_ja.jsonl
    ├── multiple_choice_classification/
    │   ├── choice2_ans1/
    │   │   ├── taxonomic_multiple-choice_classification_choice2-ans1_en.jsonl
    │   │   ├── taxonomic_multiple-choice_classification_choice2-ans1_ja.jsonl
    │   │   ├── few_shot_for_taxonomic_multiple-choice_classification_choice2-ans1_en.jsonl
    │   │   └── few_shot_for_taxonomic_multiple-choice_classification_choice2-ans1_ja.jsonl
    │   ├── choice3_ans1/
    │   │   └── ...
    │   ├── choice4_ans1/
    │   │   └── ...
    │   └── choice5_ans1/
    │       └── ...
    └── yes_no_classification/
        ├── taxonomic_yes-no_classification_en.jsonl
        ├── taxonomic_yes-no_classification_ja.jsonl
        ├── few_shot_for_taxonomic_yes-no_classification_en.jsonl
        └── few_shot_for_taxonomic_yes-no_classification_ja.jsonl

Configurations and Splits

Category	Configuration	Language	Rows (test)
Original data	original_crowdsourced_and_cleaned (default)	ja, en	22,409
Original data	original_task_a_term_selection	ja	2,697
Original data	original_task_b_term_naturalness_ja	ja	359
Original data	original_task_b_term_naturalness_en	en	349
Original data	original_no_shitsukan_terms	ja	956
Perception	perception_generation_ja	ja	107
Perception	perception_selection_{variant}_{lang}	ja, en	335 each
Commonsense	commonsense_generation_ja	ja	100
Commonsense	commonsense_classification_ja	ja	13,391
Commonsense	commonsense_classification_en	en	9,549
Taxonomic	taxonomic_yes_no_classification_ja	ja	438
Taxonomic	taxonomic_yes_no_classification_en	en	216
Taxonomic	taxonomic_a_b_classification_ja	ja	1,665
Taxonomic	taxonomic_a_b_classification_en	en	648
Taxonomic	taxonomic_multiple_choice_classification_{variant}_ja	ja	1,665 each
Taxonomic	taxonomic_multiple_choice_classification_{variant}_en	en	648 each

• {variant} is one of choice2_ans1, choice3_ans1, choice4_ans1, and choice5_ans1 (N choices with exactly one correct answer), plus random for perception selection only.
• {lang} is ja or en.
• For every task configuration, the few-shot examples used in the prompts are available as a separate configuration named <configuration>_few_shot (e.g., perception_selection_choice2_ans1_ja_few_shot; 1–4 rows each).

Full list of configurations (57)

Configuration	File	Rows (test)
original_crowdsourced_and_cleaned (default)	data/original_data/crowdsourced_and_cleaned_shitsukan_dataset.jsonl	22,409
original_task_a_term_selection	data/original_data/data_annotated_through_taskA_shitsukan_term_selection.jsonl	2,697
original_task_b_term_naturalness_ja	data/original_data/data_annotated_through_taskB_shitsukan_term_naturalness_ja.jsonl	359
original_task_b_term_naturalness_en	data/original_data/data_annotated_through_taskB_shitsukan_term_naturalness_en.jsonl	349
original_no_shitsukan_terms	data/original_data/no_shitsukan_datasets_excluded_during_cleaning.jsonl	956
perception_generation_ja	data/perception/generation/perception_generation_ja.jsonl	107
perception_generation_ja_few_shot	data/perception/generation/few_shot_for_perception-generation_ja.jsonl	1
perception_selection_choice2_ans1_ja	data/perception/selection/choice2_ans1/perception_selection_choice2-ans1_ja.jsonl	335
perception_selection_choice2_ans1_ja_few_shot	data/perception/selection/choice2_ans1/few_shot_for_perception-selection-choice2-ans1_ja.jsonl	2
perception_selection_choice2_ans1_en	data/perception/selection/choice2_ans1/perception_selection_choice2-ans1_en.jsonl	335
perception_selection_choice2_ans1_en_few_shot	data/perception/selection/choice2_ans1/few_shot_for_perception-selection-choice2-ans1_en.jsonl	2
perception_selection_choice3_ans1_ja	data/perception/selection/choice3_ans1/perception_selection_choice3-ans1_ja.jsonl	335
perception_selection_choice3_ans1_ja_few_shot	data/perception/selection/choice3_ans1/few_shot_for_perception-selection-choice3-ans1_ja.jsonl	2
perception_selection_choice3_ans1_en	data/perception/selection/choice3_ans1/perception_selection_choice3-ans1_en.jsonl	335
perception_selection_choice3_ans1_en_few_shot	data/perception/selection/choice3_ans1/few_shot_for_perception-selection-choice3-ans1_en.jsonl	2
perception_selection_choice4_ans1_ja	data/perception/selection/choice4_ans1/perception_selection_choice4-ans1_ja.jsonl	335
perception_selection_choice4_ans1_ja_few_shot	data/perception/selection/choice4_ans1/few_shot_for_perception-selection-choice4-ans1_ja.jsonl	2
perception_selection_choice4_ans1_en	data/perception/selection/choice4_ans1/perception_selection_choice4-ans1_en.jsonl	335
perception_selection_choice4_ans1_en_few_shot	data/perception/selection/choice4_ans1/few_shot_for_perception-selection-choice4-ans1_en.jsonl	2
perception_selection_choice5_ans1_ja	data/perception/selection/choice5_ans1/perception_selection_choice5-ans1_ja.jsonl	335
perception_selection_choice5_ans1_ja_few_shot	data/perception/selection/choice5_ans1/few_shot_for_perception-selection-choice5-ans1_ja.jsonl	2
perception_selection_choice5_ans1_en	data/perception/selection/choice5_ans1/perception_selection_choice5-ans1_en.jsonl	335
perception_selection_choice5_ans1_en_few_shot	data/perception/selection/choice5_ans1/few_shot_for_perception-selection-choice5-ans1_en.jsonl	2
perception_selection_random_ja	data/perception/selection/random/perception_selection_random_ja.jsonl	335
perception_selection_random_ja_few_shot	data/perception/selection/random/few_shot_for_perception-selection-random_ja.jsonl	2
perception_selection_random_en	data/perception/selection/random/perception_selection_random_en.jsonl	335
perception_selection_random_en_few_shot	data/perception/selection/random/few_shot_for_perception-selection-random_en.jsonl	2
commonsense_generation_ja	data/commonsense/generation/commonsense_generation_ja.jsonl	100
commonsense_generation_ja_few_shot	data/commonsense/generation/few_shot_for_commonsense-generation_ja.jsonl	1
commonsense_classification_ja	data/commonsense/classification/commonsense_classification_ja.jsonl	13,391
commonsense_classification_ja_few_shot	data/commonsense/classification/few_shot_for_commonsense-classification_ja.jsonl	4
commonsense_classification_en	data/commonsense/classification/commonsense_classification_en.jsonl	9,549
commonsense_classification_en_few_shot	data/commonsense/classification/few_shot_for_commonsense-classification_en.jsonl	4
taxonomic_yes_no_classification_ja	data/taxonomic/yes_no_classification/taxonomic_yes-no_classification_ja.jsonl	438
taxonomic_yes_no_classification_ja_few_shot	data/taxonomic/yes_no_classification/few_shot_for_taxonomic_yes-no_classification_ja.jsonl	2
taxonomic_yes_no_classification_en	data/taxonomic/yes_no_classification/taxonomic_yes-no_classification_en.jsonl	216
taxonomic_yes_no_classification_en_few_shot	data/taxonomic/yes_no_classification/few_shot_for_taxonomic_yes-no_classification_en.jsonl	2
taxonomic_a_b_classification_ja	data/taxonomic/a_b_classification/taxonomic_a-b_classification_ja.jsonl	1,665
taxonomic_a_b_classification_ja_few_shot	data/taxonomic/a_b_classification/few_shot_for_taxonomic_a-b_classification_ja.jsonl	2
taxonomic_a_b_classification_en	data/taxonomic/a_b_classification/taxonomic_a-b_classification_en.jsonl	648
taxonomic_a_b_classification_en_few_shot	data/taxonomic/a_b_classification/few_shot_for_taxonomic_a-b_classification_en.jsonl	2
taxonomic_multiple_choice_classification_choice2_ans1_ja	data/taxonomic/multiple_choice_classification/choice2_ans1/taxonomic_multiple-choice_classification_choice2-ans1_ja.jsonl	1,665
taxonomic_multiple_choice_classification_choice2_ans1_ja_few_shot	data/taxonomic/multiple_choice_classification/choice2_ans1/few_shot_for_taxonomic_multiple-choice_classification_choice2-ans1_ja.jsonl	2
taxonomic_multiple_choice_classification_choice2_ans1_en	data/taxonomic/multiple_choice_classification/choice2_ans1/taxonomic_multiple-choice_classification_choice2-ans1_en.jsonl	648
taxonomic_multiple_choice_classification_choice2_ans1_en_few_shot	data/taxonomic/multiple_choice_classification/choice2_ans1/few_shot_for_taxonomic_multiple-choice_classification_choice2-ans1_en.jsonl	2
taxonomic_multiple_choice_classification_choice3_ans1_ja	data/taxonomic/multiple_choice_classification/choice3_ans1/taxonomic_multiple-choice_classification_choice3-ans1_ja.jsonl	1,665
taxonomic_multiple_choice_classification_choice3_ans1_ja_few_shot	data/taxonomic/multiple_choice_classification/choice3_ans1/few_shot_for_taxonomic_multiple-choice_classification_choice3-ans1_ja.jsonl	2
taxonomic_multiple_choice_classification_choice3_ans1_en	data/taxonomic/multiple_choice_classification/choice3_ans1/taxonomic_multiple-choice_classification_choice3-ans1_en.jsonl	648
taxonomic_multiple_choice_classification_choice3_ans1_en_few_shot	data/taxonomic/multiple_choice_classification/choice3_ans1/few_shot_for_taxonomic_multiple-choice_classification_choice3-ans1_en.jsonl	2
taxonomic_multiple_choice_classification_choice4_ans1_ja	data/taxonomic/multiple_choice_classification/choice4_ans1/taxonomic_multiple-choice_classification_choice4-ans1_ja.jsonl	1,665
taxonomic_multiple_choice_classification_choice4_ans1_ja_few_shot	data/taxonomic/multiple_choice_classification/choice4_ans1/few_shot_for_taxonomic_multiple-choice_classification_choice4-ans1_ja.jsonl	2
taxonomic_multiple_choice_classification_choice4_ans1_en	data/taxonomic/multiple_choice_classification/choice4_ans1/taxonomic_multiple-choice_classification_choice4-ans1_en.jsonl	648
taxonomic_multiple_choice_classification_choice4_ans1_en_few_shot	data/taxonomic/multiple_choice_classification/choice4_ans1/few_shot_for_taxonomic_multiple-choice_classification_choice4-ans1_en.jsonl	2
taxonomic_multiple_choice_classification_choice5_ans1_ja	data/taxonomic/multiple_choice_classification/choice5_ans1/taxonomic_multiple-choice_classification_choice5-ans1_ja.jsonl	1,665
taxonomic_multiple_choice_classification_choice5_ans1_ja_few_shot	data/taxonomic/multiple_choice_classification/choice5_ans1/few_shot_for_taxonomic_multiple-choice_classification_choice5-ans1_ja.jsonl	2
taxonomic_multiple_choice_classification_choice5_ans1_en	data/taxonomic/multiple_choice_classification/choice5_ans1/taxonomic_multiple-choice_classification_choice5-ans1_en.jsonl	648
taxonomic_multiple_choice_classification_choice5_ans1_en_few_shot	data/taxonomic/multiple_choice_classification/choice5_ans1/few_shot_for_taxonomic_multiple-choice_classification_choice5-ans1_en.jsonl	2

Original Data

Contains the original human-annotated datasets:

• crowdsourced_and_cleaned_shitsukan_dataset.jsonl: The primary dataset containing shitsukan terms collected from individuals. Each row is a shitsukan term recalled for an object in a COCO image, together with its English translation.
• data_annotated_through_taskA_shitsukan_term_selection.jsonl: Data annotated through Task A (shitsukan term selection). For each image–object pair, candidate terms are grouped by the number of workers (0–5 out of 5) who selected them as suitable.
• data_annotated_through_taskB_shitsukan_term_naturalness_ja.jsonl: Data annotated through Task B (shitsukan term naturalness) in Japanese. For each object, terms are grouped by their naturalness label (1–5), i.e., the majority vote of five workers.
• data_annotated_through_taskB_shitsukan_term_naturalness_en.jsonl: English version of the Task B data.
• no_shitsukan_datasets_excluded_during_cleaning.jsonl: A list of 956 expressions that were identified as non-shitsukan terms during data cleaning (a list of terms, not of datasets).

Perception

Perception tasks evaluate whether a model perceives the shitsukan of an object in an image in the same way as humans:

• Generation:
  • perception_generation_ja.jsonl: Perception generation task in Japanese. Given an image and an object, the model generates shitsukan terms; shitsukan_list contains the terms recalled by workers for the image–object pair.
  • few_shot_for_perception-generation_ja.jsonl: Few-shot example for perception generation (text only).
• Selection:
  • choice2_ans1, choice3_ans1, choice4_ans1, choice5_ans1: Select the correct term from N (= 2–5) choices. The correct term is one selected by at least 4 of the 5 workers in Task A, and the distractors are terms selected by none of them.
  • random: An additional setting that is not reported in the paper. The number of choices varies from 2 to 5, and some items (157 of 335) have no correct choice; the expected answer for such items is -1.
  • Each subdirectory contains the data and few-shot examples in both Japanese and English.

Commonsense

Commonsense tasks evaluate knowledge of the shitsukan typically associated with an object, without images:

• Generation:
  • commonsense_generation_ja.jsonl: Commonsense generation task in Japanese. Given an object name, the model generates typical shitsukan terms; shitsukan_list contains the typical terms listed by workers.
  • few_shot_for_commonsense-generation_ja.jsonl: Few-shot example for commonsense generation.
• Classification:
  • commonsense_classification_ja.jsonl: Commonsense classification task in Japanese, built from the Task B annotations. The model judges the naturalness of an object–term pair on a 5-point scale; the gold label is the majority vote of the five workers.
  • commonsense_classification_en.jsonl: English version of the commonsense classification task (not reported in the paper).
  • few_shot_for_commonsense-classification_ja.jsonl / few_shot_for_commonsense-classification_en.jsonl: Few-shot examples for commonsense classification in Japanese / English.

Taxonomic

Taxonomic tasks evaluate whether a model can distinguish shitsukan terms from other expressions (text only). Following the paper (Section 6.1), the data are built from the naturalness annotations of commonsense_classification_ja.jsonl:

• Positive terms: terms labeled "Completely natural" or "Typical, but not necessarily" for at least one object (1,665 Japanese terms after excluding the terms used in the few-shot examples).

• Negative terms: terms for which at least one of the five workers chose "Other: not shitsukan", excluding all terms that satisfy the positive definition (219 Japanese terms). This corresponds to the harder setting of the paper (1b / 2b); the easier setting with negative terms sampled from Wikipedia (1a / 2a) is not included.

• Terms that appear in the few-shot examples are excluded from the test items.

• English: the Japanese positive and negative terms are translated with their most frequent English translation in crowdsourced_and_cleaned_shitsukan_dataset.jsonl, and negative terms whose translation coincides with a positive term are removed (648 positive and 108 negative English terms).

• The items are sampled with a fixed random seed. Because the sets are rebuilt from the released data following the definitions in the paper, they are not identical to the evaluation sets used in the paper, and scores are not directly comparable to the numbers reported there.

• Yes-No Classification:

  • taxonomic_yes-no_classification_{ja,en}.jsonl: Decide whether a given term is a shitsukan term. The data contain all negative terms and the same number of randomly sampled positive terms (ja: 219 "yes" / 219 "no"; en: 108 / 108).
  • few_shot_for_taxonomic_yes-no_classification_{ja,en}.jsonl: Few-shot examples for Yes-No classification.

• A-B Classification:

  • taxonomic_a-b_classification_{ja,en}.jsonl: Choose the shitsukan term from two candidates labeled A and B. There is one item per positive term, paired with a randomly sampled negative term, and the position of the correct answer is balanced between A and B. The items are the same as in multiple_choice_classification/choice2_ans1, with letter labels instead of numbers.
  • few_shot_for_taxonomic_a-b_classification_{ja,en}.jsonl: Few-shot examples for A-B classification.

• Multiple Choice Classification:

  • choice2_ans1, choice3_ans1, choice4_ans1, choice5_ans1: Choose the shitsukan term from N (= 2–5) candidates, exactly one of which is correct. There is one item per positive term; the distractors are nested (the N-choice item adds one negative term to the (N−1)-choice item for the same positive term), and the position of the correct choice is balanced. Each subdirectory also contains few-shot examples.

• The English few-shot examples are translations of the Japanese ones.

Data Fields

Image fields (records that refer to COCO images)

• image_id (int): COCO image ID.
• image (str; list[str] in the perception task files): relative image path, e.g., images/coco/train2017/000000241620.jpg (train2017 or val2017).
• image_url (str; list[str] in the perception task files): COCO image URL.
• flickr_url (str), height (int), width (int), license (int): COCO image metadata (license is the COCO license ID of the image).
• supercategories (list[str]), names (list[str]), category_ids (list[int]): COCO category metadata associated with the image.

Original data

• crowdsourced_and_cleaned_shitsukan_dataset.jsonl: id, object_ja / object_en (target object), shitsukan_ja / shitsukan_en (shitsukan term and its English translation; shitsukan_en is null in 5 rows), and the image fields.
• data_annotated_through_taskA_shitsukan_term_selection.jsonl: id, object, 5_vote, 4_vote, …, 0_vote (list[str]: terms selected by 5, 4, …, 0 of the 5 workers), and the image fields.
• data_annotated_through_taskB_shitsukan_term_naturalness_{ja,en}.jsonl: id, object, "1"–"5" (list[str]: terms whose naturalness label for the object is 1–5; see the label table below).
• no_shitsukan_datasets_excluded_during_cleaning.jsonl: id, no_shitsukan (str: term).

Perception

• perception_generation_ja: id, object, shitsukan_list (list[str]: reference terms), and the image fields.
• perception_selection_*: id, object, all_ids_with_shitsukan (list[str]: choices in the form "index:term"), correct_ids (list[int]: index of the correct choice; empty when there is no correct choice), uncorrect_ids (list[int]: indices of the distractors), correct_shitsukan_list (list[str]: the correct term), uncorrect_shitsukan_list (list[str]: the distractor terms), correct_ids_with_shitsukan / uncorrect_ids_with_shitsukan (list[str]: the correct choice / distractors in the form "index:term"), and the image fields.

Commonsense

• commonsense_generation_ja: id, object, shitsukan_list (list[str]: reference terms).
• commonsense_classification_{ja,en}: id, object_id, object_shitsukan_id (index of the term within the object), object, shitsukan, naturalness (int: gold label), naturalness_str (str: name of the gold label), and naturalness_by_worker (list[int]: labels given by the five workers; ja only).

The naturalness labels follow the definition in the paper (Section 5.1) and use the same numbering as the choices in the prompts:

naturalness	naturalness_str	Japanese (prompt)
1	Completely natural	完全に自然
2	Typical, but not necessarily	一般的だが、他もあり得る
3	Possible, but uncommon	あり得るが一般的ではない
4	Unnatural	不自然
5	Other: not shitsukan	その他：質感ではない

Taxonomic

• taxonomic_yes_no_classification_{ja,en}: id, shitsukan (str), label (int: 1 = shitsukan term, 0 = not a shitsukan term), label_str ("yes" / "no").
• taxonomic_a_b_classification_{ja,en}, taxonomic_multiple_choice_classification_*: id, all_ids_with_shitsukan (list[str]: choices in the form "A:term" / "0:term"), correct_ids (list[int]: index of the correct choice; for A-B classification, 0 = A and 1 = B), uncorrect_ids (list[int]), correct_shitsukan_list (list[str]: the correct term), uncorrect_shitsukan_list (list[str]: the distractors), correct_ids_with_shitsukan / uncorrect_ids_with_shitsukan (list[str]).

Few-shot examples (*_few_shot)

• user (str): input of the example, assistant (str): expected output, images (null): no image is attached to the few-shot examples.

Notes

• Majority vote (commonsense classification, ja): The gold label is the most frequent of the five workers' labels. When several labels are tied, the label in the original annotation is kept if it is one of the tied labels; otherwise, the tied label closest to the mean of the five labels is used (the smaller label if equidistant).
• commonsense_classification_ja contains 13,391 of the 13,392 object–term pairs of the Task B data; the pair of 石 and "ひんやり感　ひんやり感" (a duplicated term) is not included.
• commonsense_classification_en has no per-worker labels, and some object–term pairs appear more than once because different Japanese terms received the same English translation.

Update History

• 2026-10
  • All data are published as the test split.
  • Commonsense classification: naturalness_str now follows the label definition of the paper (the strings for labels 2 and 3 were swapped). The gold labels of 169 tied pairs that were not one of the most frequent labels were corrected to a majority label, and the Task B data (data_annotated_through_taskB_shitsukan_term_naturalness_ja.jsonl) were updated accordingly.
  • Perception selection: correct_shitsukan_list now contains the correct term, and the distractor terms are stored in the new field uncorrect_shitsukan_list (previously, correct_shitsukan_list contained the distractor terms).
  • Commonsense generation: shitsukan_list is now a JSON list (previously, a string in NumPy array notation).
  • Taxonomic: the data were rebuilt following the definitions in the paper (previously, the positive terms were all the terms of crowdsourced_and_cleaned_shitsukan_dataset.jsonl, and the negative terms were those of no_shitsukan_datasets_excluded_during_cleaning.jsonl), and English data were added. The Japanese A-B few-shot examples were replaced, because both choices of the previous examples were shitsukan terms.

Images

The images are not included in this repository. Records refer to MS COCO 2017 images (train2017 and val2017) via the image, image_url, and flickr_url fields. To run the perception tasks, please download the COCO 2017 train/val images from the COCO website.

Licensing Information

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).
The COCO images referenced by this dataset are not redistributed in this repository; they are subject to the COCO Terms of Use and the licenses of the individual images (see the license field).

Citation Information

If you use this dataset in your research, please cite:

@inproceedings{shiono-etal-2025-evaluating,
    title = "Evaluating Model Alignment with Human Perception: A Study on Shitsukan in {LLM}s and {LVLM}s",
    author = "Shiono, Daiki  and
      Brassard, Ana  and
      Ishizuki, Yukiko  and
      Suzuki, Jun",
    editor = "Rambow, Owen  and
      Wanner, Leo  and
      Apidianaki, Marianna  and
      Al-Khalifa, Hend  and
      Eugenio, Barbara Di  and
      Schockaert, Steven",
    booktitle = "Proceedings of the 31st International Conference on Computational Linguistics",
    month = jan,
    year = "2025",
    address = "Abu Dhabi, UAE",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.coling-main.757/",
    pages = "11428--11444",
}