EunsuKim/SCRIPTS

Are they lovers or friends? Evaluating LLMs' Social Reasoning in English and Korean Dialogues

Official dataset for Are they lovers or friends? Evaluating LLMs' Social Reasoning in English and Korean Dialogues.

Dataset Description

SCRIPTS is a bilingual dialogue dataset for evaluating social reasoning capabilities of Large Language Models. The dataset contains dialogues with rich annotations about relationships, social dimensions, and demographic attributes.

Dataset Splits

• en: 580 English dialogues
• ko: 567 Korean dialogues

Languages

• English
• Korean (한국어)

Dataset Structure

Data Fields

Each example contains the following fields:

Core Fields

• scene_id (string): Unique identifier for each dialogue
• dialogue (string): Conversation text with speaker markers [A]: and [B]:

Social Relation

• relation_high_probable_gold (string): Ground truth high-probability social relation
• relation_impossible_gold (string): Relations annotated as impossible/unlikely
• high_probable_agreement (string): Inter-annotator agreement level

Social Dimensions

• intimacy_gold (string): Intimacy level (intimate, not intimate, neutral, unknown)
• intimacy_agreement (float): Inter-annotator agreement score
• formality_gold (string): Formality/task orientation (formal, informal, neutral, unknown)
• formality_agreement (float): Inter-annotator agreement score
• hierarchy_gold (string): Power dynamics (equal, hierarchical, unknown)
• hierarchy_agreement (float): Inter-annotator agreement score

Demographics

• age-a_gold (string): Age category for speaker A
• age-b_gold (string): Age category for speaker B
• age_diff_gold (string): Age comparison (A>B, A<B, A=B, Unknown)
• gender-a_gold (string): Gender for speaker A
• gender-b_gold (string): Gender for speaker B
• gender_diff_gold (string): Gender comparison (Same, Different, Unknown)

Data Example

English (en split):

{
  'scene_id': 'scene300',
  'dialogue': '[B]: [A], right? Happy to meet you.\n[A]: Officially almost human again...',
  'relation_high_probable_gold': "{'rank1': {'Police-Victim': 0.67, 'Police officer-Civilian': 0.33}, ...}",
  'intimacy_gold': 'Unintimate',
  'formality_gold': 'Task-oriented',
  'hierarchy_gold': 'A<B',
  'age-a_gold': "['(20–35) Young adult']",
  'gender-a_gold': "['Cannot be determined']",
  ...
}

Korean (ko split):

{
  'scene_id': '0',
  'dialogue': 'B: 눈깔 안 돌리면 뽑아서 골프공으로 쓴다!...고 속으로 말했습니다...',
  'relation_high_probable_gold': "['친구', '성직자-신도', '지인']",
  'intimacy_gold': '친함',
  'formality_gold': '즐거움 중심',
  'hierarchy_gold': 'A=B',
  'age-a_gold': "['대학생(20-24)', '청년(25-39)', '중장년(40-59)', '노년(65-)']",
  'gender-a_gold': "['남성', '여성']",
  ...
}

Usage

Loading the Dataset

from datasets import load_dataset

# Load both splits
dataset = load_dataset('EunsuKim/SCRIPTS')

# Access English dialogues
en_data = dataset['en']
print(f"English samples: {len(en_data)}")

# Access Korean dialogues
ko_data = dataset['ko']
print(f"Korean samples: {len(ko_data)}")

# View a sample
print(en_data[0])

Loading Specific Split

# Load only English
en_dataset = load_dataset('EunsuKim/SCRIPTS', split='en')

# Load only Korean
ko_dataset = load_dataset('EunsuKim/SCRIPTS', split='ko')

Example: Filtering by Relation Type

from datasets import load_dataset

dataset = load_dataset('EunsuKim/SCRIPTS', split='en')

# Filter dialogues with high intimacy
intimate_dialogues = dataset.filter(lambda x: 'intimate' in x['intimacy_gold'].lower())
print(f"Found {len(intimate_dialogues)} intimate dialogues")

License

CC-BY-NC-ND 4.0 (Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International)

Citation

@misc{kim2025loversfriendsevaluatingllms,
      title={Are they lovers or friends? Evaluating LLMs' Social Reasoning in English and Korean Dialogues}, 
      author={Eunsu Kim and Junyeong Park and Juhyun Oh and Kiwoong Park and Seyoung Song and A. Seza Dogruoz and Najoung Kim and Alice Oh},
      year={2025},
      eprint={2510.19028},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.19028}, 
}

Contact

For questions or issues, please:

• Open an issue on GitHub
• Refer to the paper for detailed methodology