sy-volkov/hk-synthetic-personas

Hong Kong Synthetic Personas for Campaign Survey Simulation

This dataset contains synthetic Hong Kong resident personas for campaign survey simulation and message pre-testing with language models.

It includes:

• hk_personas_1000.csv and hk_personas_1000.jsonl: a broad synthetic pool of 1,000 personas.
• hk_personas_100.csv and hk_personas_100.jsonl: a curated 100-person panel selected from the 1,000-person pool.
• docs/: source notes, methodology, group definitions, validation report, and a reusable survey prompt template.
• scripts/: the deterministic generator and helper scripts used to create prompts and summarize survey results.
• data/raw/: compact official Hong Kong 2021 Population Census source snapshots used for reproducibility.
• data/processed/: generated distribution summaries for the pool and curated panel.

Intended Use

Use this dataset to explore possible campaign reactions, objections, message risks, channel differences, language accessibility issues, and segment-level response variation.

Do not use it to claim real Hong Kong public opinion, estimate vote share, forecast survey results, or replace real respondent research.

Data Source Basis

The main source is the Hong Kong 2021 Population Census District Profiles CSV for District Council Districts, published by the Census and Statistics Department.

The generator extracts aggregate distributions for:

• district and region;
• age and sex;
• education;
• occupation and employment status;
• household income;
• housing type;
• household composition;
• broad ethnicity fields;
• usual spoken language;
• birthplace and migration/residency background signals.

Source links and field notes are documented in docs/data_sources.md.

Method Summary

The pipeline is:

1. Parse official census district-profile tables.
2. Generate 1,000 demographic skeletons from census-weighted distributions.
3. Apply controlled oversampling for campaign-relevant smaller groups, including non-Chinese or non-Cantonese language profiles, older one-person households, lower-income public-rental residents, international or English-dominant residents, and lone-parent/caregiving household patterns.
4. Infer 12 demographic segments by clustering skeleton features rather than manually defining groups first.
5. Generate short and long persona cards one by one.
6. Validate each record for required fields, duplicate cards, obvious name/address patterns, excessive concentration, and group coverage.
7. Select a data-aware, diversity-maximizing 100-person panel from the pool.

Full methodology is in docs/methodology.md; inferred segment details are in docs/groups_definition.md.

Schema

Each persona includes:

• demographic fields such as age, age_group, gender, district, housing_type, household_composition, education_level, occupation, employment_status, income_level, language_profile, migration_or_residency_background, and ethnicity_or_community_background;
• simulation-readiness fields such as media_habits, shopping_or_consumption_style, values_and_priorities, trust_level, price_sensitivity, novelty_openness, brand_skepticism, likely_campaign_lens, possible_objections, and likely_channels_seen;
• prompt-ready narrative fields: short_persona_card and long_persona_card.

Validation

The current version passed the automated checks in docs/validation_report.md:

• 1,000 pool rows and 100 curated rows.
• No duplicate short or long persona cards.
• No obvious real-name/title or exact-street-address patterns.
• All major inferred groups covered in the curated panel.
• No single district above 12 percent of the pool.
• No occupation detail above 16 percent of the pool.
• Internal value diversity within every inferred group.

Limitations

• This is not a statistically valid sample.
• Census district profiles are aggregate marginal tables, not respondent-level microdata.
• Cross-variable combinations are synthetic approximations.
• Persona attitudes and campaign response styles are inferred for simulation readiness and are not measured survey responses.
• Controlled oversampling intentionally changes some small-group representation to improve message-risk testing coverage.
• Sensitive identity fields are broad and non-identifying by design.

Example Survey Prompt

See docs/sample_survey_prompt.md for a reusable prompt template. To generate dry-run prompt payloads:

python3 run_persona_survey.py --personas hk_personas_100.jsonl --campaign campaign.txt --question question.txt --dry-run

Recommended Validation Step

Before relying on synthetic results for decisions, compare outputs against a small real Hong Kong survey stratified by age, region/district, housing or income band, and language profile. Use the comparison to calibrate objections, confusion points, and message-risk themes rather than treating synthetic response percentages as opinion estimates.

Citation

If you use this dataset, please cite:

@dataset{volkov_2026_hk_synthetic_personas,
  author    = {Volkov, Sergey},
  title     = {HK Synthetic Personas: Census-Grounded Synthetic Personas for Hong Kong},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.0},
  doi       = {10.5281/zenodo.20922880},
  url       = {https://doi.org/10.5281/zenodo.20922880}
}