CC-Mediation

A cross-cultural conflict-mediation benchmark grounded in the Developmental Model of Intercultural Sensitivity (DMIS). Each scenario is a culturally grounded conflict dialogue with a mediation intervention and its post-intervention trajectory, organised as a preference pair (positive vs negative continuation) so downstream effects are measurable.

📄 Paper: CC-Mediation: Evaluating Large Language Models for Cross-Cultural Conflict Mediation (Lee, Zhang, Yow, Deng; arXiv:2609.04855) 💻 Code: https://github.com/suhyun565/CC-Mediation

What is in a row

One row = one scenario, with three continuation branches laid side by side so you can filter by conflict_stage or sort by delta_judge_score directly, and the original turn structure is preserved.

group	fields
scenario	scenario_id, subset, conflict_stage, conflict_turn, conflict_agent, max_turns, scenario_context
agents	agent_1_*, agent_2_* (name, nationality, occupation, age, gender), agent_{1,2}_value, agent_{1,2}_goals
dialogue	shared_prefix (turns before the branch)
branches	for b ∈ {original, positive, negative}: {b}_turns, {b}_path_label, {b}_intervention_turn, {b}_mediator_model, {b}_mediator_utterance, {b}_judge_score, {b}_trajectory_auc
preference	chosen, rejected, pair_type, mediator_model_chosen, mediator_model_rejected, delta_judge_score, delta_trajectory_auc

Each turn (shared_prefix, {b}_turns) has: t, speaker, agent, message, phase, label.

Splits & composition

split	rows
train	1,503
test	159
total	1,662 scenarios

• conflict_stage (DMIS): Denial 520 · Minimization 514 · Defense 469
• pair_type: timing-differs 1,313 · content-differs 190

Evaluation metrics (from the paper)

• Trajectory AUC — persistence of intercultural improvement over time (*_trajectory_auc, delta_trajectory_auc).
• Signed Wasserstein-1 — magnitude and direction of intercultural stance shift.
• A DMIS-grounded judge score (*_judge_score, delta_judge_score); both metrics agree strongly with human judgment.

Usage

from datasets import load_dataset
ds = load_dataset("Suhyunlee/cc-mediation")

# scenarios where mediation helped most
top = ds["train"].sort("delta_judge_score", reverse=True)[:10]

# filter by DMIS stage
denial = ds["train"].filter(lambda r: r["conflict_stage"] == "Denial")

Citation

@article{lee2026ccmediation,
  title  = {CC-Mediation: Evaluating Large Language Models for Cross-Cultural Conflict Mediation},
  author = {Lee, Suhyun and Zhang, Wenxuan and Yow, W. Quin and Deng, Yang},
  journal = {arXiv preprint arXiv:2609.04855},
  year   = {2026},
  url    = {https://arxiv.org/abs/2609.04855}
}

License

Not yet specified — set the license field in this card before making the dataset broadly available.