Version 1.0.0 is available through gated research access. The structured tables and sanitized participant-authored text pass the documented release checks. Manual review covered every high-priority passage, a fixed random sample of 600 medium-priority passages, and 155 additional medium-priority passages. Text outside the manual sample received the same automated sanitization applied to every row. Participant consent, research use, and deidentified release were covered by the company-approved data agreement used for collection.

This dataset is distributed separately from the public synthetic HorizonBench benchmark. The canonical versioned release is the gated Hugging Face dataset at https://huggingface.co/datasets/stellalisy/preference-forecast. Access is provided after manual approval.

Study Overview

The dataset contains seven surveys collected over 13 weeks from 160 participants who completed T0 through T6. Participants rated 52 recurring everyday preferences, made forecasts of future ratings, described planned events, and rated preferences associated with those events. A linked post-study survey records forecasting strategies and views about personalized AI.

HorizonBench uses only immediate event-associated rating changes to calibrate the probability and magnitude of synthetic preference evolution. The release contains additional numeric forecasts and sanitized text for separate cognitive research. Their inclusion does not make those analyses part of the HorizonBench contribution.

Quick Start

After receiving access, follow QUICKSTART.md to download the dataset and run:

python scripts/load_longitudinal_data.py --output-dir analysis_ready

The command uses only the Python standard library and produces joined rating trajectories, explicit forecast-to-outcome pairs, labeled missingness, and normalized final-survey domain selections without modifying the source files.

Files

data/participants.csv: randomized participant IDs and completion status.
data/core_items.csv: fixed study item IDs and wording.
data/ratings_long.csv: current ratings and all available forecast horizons for core and event-linked preferences.
data/event_status.csv: event completion status indexed by participant and source survey.
data/free_text_sanitized.csv: sanitized event narratives, preference wording and explanations, prediction reflections, life-event descriptions, and final-survey explanations.
data/final_survey_structured.csv: structured linked post-study survey responses.
data/demographics_aggregate.csv: aggregate demographics with cells smaller than five suppressed.
LONGITUDINAL_INSTRUMENT.md: verified survey schedule, protocol wording, response options, and implemented attention checks.
LONGITUDINAL_FIELD_MAP.csv: mapping from released longitudinal columns to surveys, meanings, response formats, and source fields.
LONGITUDINAL_PROTOCOL_VALIDATION.md: confirmed implementation details, differences from the planning protocol, and remaining wording uncertainty.
FINAL_SURVEY_INSTRUMENT.md: exact revised survey wording, response options, scale anchors, and released variable names.
metadata/participant_flow.json: inclusion, exclusion, and sanitization summary.
metadata/release_manifest.json: file sizes and SHA-256 checksums.
scripts/reproduce_horizonbench_calibration.py: reproduces the HorizonBench calibration estimates from the released numeric tables.
scripts/load_longitudinal_data.py: creates joined, analysis-ready longitudinal tables and labels structural blanks separately from missing responses.
scripts/audit_release_package.py: checks the assembled package for risky columns, direct-identifier patterns, and incomplete text review.
scripts/update_release_manifest.py: regenerates checksums after an approved release is assembled.

Start with QUICKSTART.md. See LONGITUDINAL_INSTRUMENT.md, LONGITUDINAL_PROTOCOL_VALIDATION.md, LONGITUDINAL_FIELD_MAP.csv, DATA_DICTIONARY.md, FINAL_SURVEY_INSTRUMENT.md, SANITIZATION_PROTOCOL.md, and MANUAL_TEXT_REVIEW.md for field definitions, survey wording, implementation checks, and privacy controls. Row-level review flags, priorities, and review statuses remain in the restricted working repository and are not distributed.

Access

Applicants must submit the information in ACCESS_REQUEST_TEMPLATE.md and execute the approved data use agreement. Access is limited to noncommercial research. Reidentification, participant contact, redistribution, and attempts to link these records to external personal data are prohibited.

Consent and Governance

All participants completed a data agreement covering research use and release of deidentified responses. Data collection and release are governed through company legal and privacy review. The release contains no names, emails, vendor identifiers, original participant IDs, row-level demographics, exact dates, or unsanitized text. See CONSENT_AND_GOVERNANCE.md for the release statement and STUDY_PROTOCOL.md for the study design.

Known Limits

The cohort is concentrated among adults aged 18 to 39, urban residents, bachelor's-degree holders, and people working in data annotation or AI training. The sample should not be treated as demographically representative of a national population. See SAMPLE_CHARACTERISTICS.md.

Citation

Dataset: https://huggingface.co/datasets/stellalisy/preference-forecast

The dataset should be cited through the HorizonBench paper, which introduces the collection and release.

@misc{li2026horizonbench,
  title = {HorizonBench: Long-Horizon Personalization with Evolving Preferences},
  author = {Li, Shuyue Stella and Paranjape, Bhargavi and Oktar, Kerem and Ma, Zhongyao and Zhou, Gelin and Guan, Lin and Zhang, Na and Park, Sem and Chen, Lin and Yang, Diyi and Tsvetkov, Yulia and Celikyilmaz, Asli},
  year = {2026},
  eprint = {2604.17283},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2604.17283}
}

Downloads last month: 37

Total file size:

9.38 MB

Paper for stellalisy/preference-forecast

HorizonBench: Long-Horizon Personalization with Evolving Preferences

Paper • 2604.17283 • Published Apr 19 • 1