CAREBench Data Documentation

1. Overview

CAREBench is a diagnostic benchmark for evaluating LLMs’ emotion understanding through cognitive appraisal reasoning. Its annotations cover four connected stages—situational narratives, open-ended appraisal reasoning, appraisal ratings, and emotion outcomes—enabling process-level evaluation along the appraisal–emotion chain. By combining first-person self-reports with independent third-person interpretations, CAREBench also supports the study of perspective differences and subjective variation in emotion understanding.

The dataset provides:

• Dataset composition: 1,000 first-person narratives with self-annotations, including a matched subset of 500 narratives with five independent third-person annotations each.
• Appraisal annotations: open-ended reasoning on five core appraisal dimensions and ratings on 22 fine-grained appraisal dimensions.
• Emotion annotations: positive and negative emotion intensity ratings and multi-label emotion categories.
• Annotator attributes: BFI-10 responses, computed Big Five scores, and demographic information.

2. Dataset Organization

CAREBench is distributed as two JSON Lines files:

data/
├── first_person.jsonl
└── third_person.jsonl

File	Contents
first_person.jsonl	1,000 first-person narratives and their self-annotations, with one record per line.
third_person.jsonl	Five independent observer annotations for each of 500 narratives selected from first_person.jsonl, with one observer record per line.

2.1 First-Person Data

Each line in first_person.jsonl is one first-person annotation record:

{"sample_id":"...","participant_id":"...","story_collection":{},"cognitive_questions":{},"appraisal_ratings":{},"emotion_labels":{},"personality_questionnaire":{},"BigFive_scores":{},"demographics":{}}

Each sample ID identifies one narrative and its associated self-annotations.

2.2 Third-Person Data

Each line in third_person.jsonl is one observer annotation record:

{"sample_id":"...","observer_index":1,"participant_id":"...","cognitive_questions":{},"appraisal_ratings":{},"emotion_labels":{},"personality_questionnaire":{},"BigFive_scores":{},"demographics":{}}

The same sample_id therefore appears on five lines, with observer_index ranging from 1 to 5.

2.3 Linking the Two Files

The sample IDs in third_person.jsonl form a matched subset of the sample IDs in first_person.jsonl.

For a given sample ID:

• first_person.jsonl provides the source narrative and the experiencer’s self-annotations;
• third_person.jsonl provides five independent observer interpretations of that narrative.

Observer records do not duplicate the story_collection object. The corresponding source narrative should be retrieved from the record with the same sample ID in first_person.jsonl.

3. Annotation Record Schema

First-person and third-person records share the same annotation schema except for story_collection, which appears only in first-person records.

Field	Type	First person	Third person	Description
participant_id	string	Yes	Yes	Identifier associated with the annotator.
story_collection	object	Yes	No	Story-elicitation dialogue, structured answers, and finalized narrative.
cognitive_questions	object	Yes	Yes	Open-ended appraisal-reasoning annotations.
appraisal_ratings	object	Yes	Yes	Ratings for 22 appraisal dimensions and one attention-check item.
emotion_labels	object	Yes	Yes	Positive and negative emotion intensities and categorical labels.
personality_questionnaire	object	Yes	Yes	Raw responses to the ten BFI-10 items.
BigFive_scores	object	Yes	Yes	Computed scores for the five Big Five personality dimensions.
demographics	object	Yes	Yes	Demographic attributes associated with the annotator.

In first-person records, participant_id refers to the experiencer who supplied the narrative. In third-person records, it refers to the observer who annotated the matched narrative.

4. Field Definitions

4.1 story_collection

story_collection appears only in first-person records and contains the participant-led, LLM-assisted story-collection process.

Field	Type	Description
final_scenario	string	Participant-reviewed first-person narrative used as the situation input in the experiments.
summary_answers	object	Structured answers collected before the narrative was assembled.
chat_history	array	Original story-elicitation dialogue represented as [role, text] pairs.

story_collection.summary_answers

Field	Type	Description
event	string	What happened in the event.
goal	string	What the participant wanted or needed in the situation.
feeling	string	How the situation left the participant feeling.
context	string	Additional context considered important by the participant.

story_collection.chat_history

Each dialogue turn is stored as a two-element array:

["ai", "Interviewer question"]

or:

["human", "Participant response"]

The role ai identifies the structured interviewer, while human identifies the participant.

4.2 cognitive_questions

cognitive_questions contains the open-ended appraisal-reasoning annotations.

Field	Type	Description
final_scenario	string	Finalized natural-language appraisal-reasoning text assembled from the dimension-specific answers.
summary_answers	object	Open-ended answers for the five core appraisal dimensions.
chat_history	array	Original appraisal interview represented as [role, text] pairs.

cognitive_questions.summary_answers

Field	Type	Description
previousStory	string	Narrative presented as context for the appraisal questions.
relevance	string	Why and to what extent the event mattered.
certainty	string	Whether the event unfolded as expected and what was unexpected.
congruence	string	How the event helped or harmed the annotator.
control	string	Whether and why the annotator felt in control.
accountability	string	Who or what was considered responsible for the event.

The chat_history entries use the same [role, text] representation as story_collection.chat_history.

4.3 appraisal_ratings

appraisal_ratings is an object whose keys are the appraisal statements presented to the annotator and whose values are five-point Likert-scale responses.

It contains:

• 22 appraisal-rating fields used in the benchmark; and
• one attention_check field used during data-quality screening.

The attention-check field is not one of the 22 appraisal dimensions and should be excluded from appraisal evaluation.

4.4 emotion_labels

Despite its name, emotion_labels contains both emotion-intensity ratings and categorical emotion labels.

Field	Type	Description
positive_level	string	Positive-emotion intensity on a 0–6 scale.
negative_level	string	Negative-emotion intensity on a 0–6 scale.
positive_emotion_labels	array of strings	All selected positive emotion groups.
negative_emotion_labels	array of strings	All selected negative emotion groups.

Multiple labels may be selected for the same event. Positive and negative emotion labels are not mutually exclusive.

4.5 personality_questionnaire

personality_questionnaire stores the raw responses to the ten BFI-10 statements. Each object key is a questionnaire statement, and its value is the annotator’s Likert-scale response.

The ten fields are:

• I see myself as someone who is reserved.
• I see myself as someone who is generally trusting.
• I see myself as someone who tends to be lazy.
• I see myself as someone who is relaxed, handles stress well.
• I see myself as someone who has few artistic interests.
• I see myself as someone who is outgoing, sociable.
• I see myself as someone who tends to find fault with others.
• I see myself as someone who does a thorough job.
• I see myself as someone who gets nervous easily.
• I see myself as someone who has an active imagination.

4.6 BigFive_scores

BigFive_scores contains the computed personality scores derived from the BFI-10 responses.

Field	Type	Range in the released data
Extraversion	number	0–10
Agreeableness	number	0–10
Conscientiousness	number	0–10
Neuroticism	number	0–10
Openness	number	0–10

4.7 demographics

demographics contains the following string-valued fields:

Field	Type
Age	string
Sex	string
Ethnicity simplified	string
Country of birth	string
Country of residence	string
Nationality	string
Language	string
Student status	string
Employment status	string

5. Value Scales and Label Sets

5.1 Appraisal-Rating Scale

The 22 appraisal dimensions use the following five-point scale:

Stored response	Numeric value used for evaluation
Strongly disagree	1
Somewhat disagree	2
Neither agree nor disagree	3
Somewhat agree	4
Strongly agree	5

5.2 Appraisal Dimensions

Dimension	Statement stored in appraisal_ratings
relevance.general	This situation matters to me.
relevance.urgency	I need to do something about this situation right away.
relevance.goals	This situation relates to my goals and plans.
relevance.bodily motives	This situation relates to my physical well-being.
relevance.social motives	This situation involves people who matter to me.
relevance.identity motives	This situation concerns who I am and what I stand for.
certainty.construal	It is clear to me what is going on in this situation.
certainty.outlook	I know what will come next in this situation.
certainty.predictability	I saw this situation coming.
certainty.novelty	This is a new kind of situation for me.
congruence.general	This is a good situation.
congruence.outlook	This situation will get better with time.
congruence.positive prediction error	This situation is better than I expected.
congruence.negative prediction error	This situation is worse than I expected.
control.general	This situation is under my control.
control.select	I can decide whether to stay in this situation or leave it.
control.vicarious	Someone can handle this situation for me.
control.effortful	I have to exert effort in this situation.
accountability.self	I am responsible for this situation.
accountability.other	Someone else is responsible for this situation.
accountability.intentionality	This situation was caused intentionally.
accountability.fairness	This situation is fair and deserved.

5.3 Emotion-Intensity Scale

The stored intensity strings include both the numeric value and its verbal description.

Value	Positive intensity	Negative intensity
0	Not at all positive	Not at all negative
1	Very slightly positive	Very slightly negative
2	Slightly positive	Slightly negative
3	Moderately positive	Moderately negative
4	Quite a bit positive	Quite a bit negative
5	Very positive	Very negative
6	Extremely positive	Extremely negative

Example:

{
  "positive_level": "2 - Slightly positive",
  "negative_level": "5 - Very negative"
}

5.4 Positive Emotion Labels

The predefined positive emotion groups are:

• Hopeful, optimistic, encouraged
• Grateful, appreciative, thankful
• Glad, happy, joyful
• Love, closeness, trust
• Amused, entertained
• Sympathy, concern, compassion
• Inspiration, admiration, appreciation
• Calm, peaceful, relaxed
• Proud, confident, determined
• Excited, enthusiastic, elated

5.5 Negative Emotion Labels

The predefined negative emotion groups are:

• Angry, frustrated, annoyed
• Worried, nervous, fearful
• Sad, downhearted, unhappy
• Disgust, distaste, revulsion
• Despair, hopelessness, sorrow
• Ashamed, humiliated, embarrassed
• Lonely, isolated, disconnected from others
• Panicked, alarmed, freaked out
• Guilty, blameworthy, repentant
• Confused, disoriented, surprised

An emotion-label array may be empty when no label from the corresponding category was selected.

5.6 BFI-10 Response Scale

Responses appearing in personality_questionnaire include:

• Strongly disagree
• Disagree
• Neither agree nor disagree
• Somewhat agree
• Strongly agree

6. License

The dataset is released under the CC BY-NC-ND 4.0 license.