InterSocialBench: human and LLM response data

This dataset accompanies InterSocialBench: Benchmarking Human and LLM Preferences for Companion-Robot Social Behavior, by Yaodan Xu, Boyang Guo, Yuqing Gu, Qingxin Zhang, Yiwen Deng, Meng Liu, and Lintian Li (Xdream Robotics). Yaodan Xu and Boyang Guo contributed equally. Correspondence: yaodanx@gmail.com and lynk@xdreamrobo.com.

What is included

The release contains the prepared formal-question data used by the benchmark paper, original Chinese scenario and candidate text, a frozen evaluation manifest, the complete author-confirmed AI candidate export including reasons, and executable analysis code. It does not contain the separate personalization or robot-execution study.

Component	Coverage
Scenarios	210 scenarios; 18 high-level behaviors; 2,145 feasible scenario–candidate pairs
Human respondents	100 pseudonymous participants; 40 distinct formal questions per person
Human observations	4,000 participant–question observations; each question answered by 12–28 participants
Human candidate judgments	41,003 rows, retaining fit, unfit, unclear/unmarked, and preferred-choice fields
Valid human preferred choices	3,830 observations; 107 lack a unique preferred choice and 63 mark the preferred action unfit
AI collection	7 models × 16 personality conditions × 210 scenarios = 23,520 response groups
AI candidate judgments	240,240 rows, including the model-generated reason field
Valid AI preferred choices	23,519 groups; one group has no preferred choice
Evaluation	Five frozen scenario-grouped folds, 42 scenarios each, 123 groups

An observation is one participant's response to one scenario. A candidate row is that participant's judgment of one feasible behavior in that scenario. Neither 41,003 nor 240,240 is a participant count. Human MBTI or demographic profiles are not included. AI personality conditions are prompting labels, not human participant attributes.

Quick start

Use Python 3.10 or later. The data use UTF-8 and preserve the original Chinese wording; the documentation is in English, with a Chinese introduction in README_zh.md.

python -m pip install -r requirements.txt
python validate.py
python reproduce.py

validate.py checks keys, candidate coverage, split integrity, response counts, and the exact relationship between the full AI export and the analysis projection. reproduce.py writes results/analysis.json. No paid API, model download, or training is required. To keep the included results untouched:

python reproduce.py --output-dir reproduced_results

Files and joins

Start with data/case_manifest.jsonl for the scenario text and ordered feasible actions. Join human case_id and AI question_id to scenario case_id; these are the same Qxxx identifiers. Join human mode_id to candidate candidate_id. AI exports use Chinese candidate names. The analysis code maps these names to behavior IDs and explicitly handles the GLANCE alias 瞄一眼后继续互动或当前任务.

• data/human_labels_deidentified.csv: all prepared formal human candidate judgments, including observations with invalid preferred choices.
• data/observation_validity.csv: observation-level validity and exclusion reasons. Apply this mask for the primary choice-based metrics; do not silently delete the underlying judgments.
• data/ai_candidate_judgments_full.csv: the complete author-confirmed new-bank export, including reasons.
• data/ai_candidate_judgments.csv: an exact projection of the full export with only reason removed; the two files are the same collection, not additional observations.
• data/question_manifest_with_candidates.jsonl: frozen folds and construction metadata. Use fold for the paper's evaluation, not export_fold.
• data/behavior_categories.csv: the paper's mapping of 18 behavior IDs to seven categories.
• data/model_inventory.csv: export IDs and author-reported model names. Blank collection metadata are unavailable, not default settings.
• DATA_DICTIONARY.md, PROTOCOL.md, and DATA_STATEMENT.md: field definitions, analysis rules, and provenance/limitations.
• reproduce.py, requirements.txt, and results/analysis.json: executable analysis and its output.
• validate.py, results/validation.json, and SHA256SUMS.json: release validation and file-integrity information.

The question manifest includes designer focal actions and other construction metadata. These are not human labels and must not be used as answer-bearing predictor inputs. The same participants may appear across folds: this protocol evaluates unseen scenario groups, not unseen people.

Reference results

The following values are percentages, pooled over 3,830 valid human observations, ordered as Top-1 / Top-3 / Accept / Reject:

Baseline	Top-1	Top-3	Accept	Reject
Uniform random	10.33	31.00	48.42	37.33
Training frequency	15.43	39.14	40.21	44.67
GPT-5.6-sol persona voting	40.97	68.22	81.46	11.67
Seven-model persona voting	39.77	71.46	82.22	11.96

Matched distinct choices per scenario: humans 4.68; models 2.06–3.46. Human plurality agreement is 51.5% when each scenario receives equal weight. This descriptive statistic is not a universal prediction ceiling.

Citation and reuse

Use the dataset record's citation or CITATION.cff. The paper and dataset are distinct research objects; do not use the paper's future arXiv identifier as the dataset DOI.

The data and documentation are licensed under CC BY 4.0, allowing reuse and adaptation, including commercial use, subject to attribution and the other license terms. The reproduction and validation code is licensed under MIT. See LICENSE.md for the exact scope and full license texts.