peach-lab/CIDER

CIDER: A Dataset of Contextual Disclosure Boundaries for Privacy Preference Alignment

Paper | Code

Dataset for the COLM 2026 paper CIDER: A Dataset of Contextual Disclosure Boundaries for Privacy Preference Alignment

CIDER is a dataset of privacy disclosure decisions collected from real users. It consists of 14,850 annotations from 169 users, forming 1,650 contextual disclosure boundary sets across 60 interpersonal communication scenarios.

What can you do with CIDER?

1. 1.Explore individual privacy disclosure preferences. Each scenario provides nine disclosure variants arranged on a granularity × identifiability grid (G1-G3 × I1-I3), and 2-3 role-specific visual cards. Each participant rate YES/NO willingness to disclose for variants of a scenario from an assigned communication role and AI-mediated condition. Each participant has at least 8 diverse scenarios to depict how their disclosure preference manifest contextually.
2. 2.Evaluate LMs with in-context learning prediction tasks. Please see our codebase for the evaluation set-up.
3. 3.Reuse or extend new study material with the toolkit. You are welcome to reuse the study toolkit for further data collection or extend the dataset by generating nine disclosure variants for new scenarios. Please see our codebase for the generation pipeline.

Dataset

data/
├── scenarios.jsonl            # 60 scenarios × 9 disclosure variants each
├── user_ratings.jsonl         # 169 participants × up to 10 scenarios × 9 variant ratings
└── visual_card_artifacts/     # 160 PNG visual cards, {scenario_id}_{role}.png

Each scenario is presented from the perspective of 2 or 3 roles (data sender, data recipient, and —— when different from the sender —— data subject), and each role perspective is crossed with a Human vs. AI-mediated framing, yielding the 320 context variations. Each participant was assigned one role and one framing condition, then rated 8-10 contexts; every context pairs one visual card with 9 disclosure variants, each receiving a binary rating. One participant's 9 ratings for one context form a contextual disclosure boundary set.

scenarios (60 rows, 39 fields)

Each row is one interpersonal communication scenario.

user_ratings (169 rows, 125 fields)

Each row is one participant's ratings with assigned role, AI condition, and de-identified demographic information.

To analyze ratings, join s{k}_id against scenario_id in the scenarios config and match var{ij} to variant_{ij}.

Visual cards

data/visual_card_artifacts/ contains 160 PNGs named {scenario_id}_{role}.png where role is data_sender, data_recipient, or data_subject (subject cards exist only for the 40 scenarios where the subject differs from the sender).

Usage

Please see our codebase for quick start guide.

Dataset Creation

Scenarios were selected from PrivacyLens seeds, and disclosure variants were generated with the pipeline in our codebase. Each participant was assigned one role, one AI-mediated condition, and a random set of scenarios presented as contexts matching that role and condition. For further study details, please refer to our paper [CIDER]().

Considerations

• —Scenarios describe sensitive disclosure situations (health, legal, relationship, and safety topics) by design; all named individuals are fictional.
• —Participant data is de-identified and limited to coarse demographics.

License

MIT. See LICENSE.

Citation

@article{guo2026cider,
  title  = {CIDER: A Dataset of Contextual Disclosure Boundaries for Privacy Preference Alignment},
  author = {Guo, Bingcan and Xu, Eryue and Zhou, Jijie and Zhang, Zhiping and Li, Tianshi},
  journal = {arXiv preprint arXiv:2608.09164}
  year   = {2026}
}