DATASHEET.md

Datasheet: Spain Reference Personas Frontier

1. Overview

Spain Reference Personas Frontier is a synthetic reference population and benchmark package for Spanish-language LLM work grounded in the territorial, household, cultural, linguistic, and civic structure of Spain. The dataset is built for controllable evaluation and simulation, not for long-form biography generation as an end in itself.

Snapshot fact	Value
Release id	spain-reference-personas-2025-v0.1
Population reference date	2025-12-31
Release date	2026-03-20
Adult personas	1,000,000
Households	536,741
Persona views	6,350,524
Actor-state rows	1,000,000
Benchmark tasks	1,800

2. Motivation

The central design decision is that a persona should be a package. Serious LLM workflows need structured controls for filtering, compact prompt views for conditioning, mutable state for event-driven simulation, and benchmark tasks for evaluation. The release therefore prioritizes usefulness, composability, and reproducibility over narrative volume.

The project takes inspiration from multi-view persona releases such as Nemotron Personas USA, but adapts that idea to Spain-specific benchmarking, household-aware simulation, and explicit held-out evaluation splits.

3. Population universe

• Universe: adult residents of Spain, age 18+.
• Scope: Spain-specific, not a generic Spanish-language population.
• Minors are represented only through household context in v0.1.
• Public views are Spanish-first.
• Co-official languages and immigrant-language repertoire are modeled as metadata.
• Public geography is limited to region and municipality class.

4. Artifact inventory

Artifact	Rows	Grain	Role
persona_core.parquet	1,000,000	person	Stable synthetic adult structure
household_core.parquet	536,741	household	Household composition and economic context
persona_views.parquet	6,350,524	person-view	Spanish LLM-facing renderings with token budgets
actor_state_init.parquet	1,000,000	person	Mutable simulation-state scaffold
benchmark_tasks.parquet	1,800	task	Tasks, splits, scoring targets, replay seeds
source_registry.parquet	11	source	Release-level source inventory
field_provenance.parquet	13	field-group	Per-field provenance mapping
EVALUATION_METRICS.json	1	release	Machine-readable evaluation summary

5. Intended uses

Workflow	Recommended artifacts	Why
Polling and policy tradeoffs	persona_core, persona_views, benchmark_tasks	structured retrieval plus policy framing and replayable task scoring
Consumer and inflation studies	household_core, persona_core, consumer_view	household burden and purchase constraints materially affect prompts
Media and event-reaction simulation	actor_state_init, dialogue_view, culture_view	mutable recency and media exposure matter for scenario responses
Sociological subgroup design	persona_core, household_core	region, language, migration, household form, and values remain queryable
Multi-turn agent benchmarks	all configs	stable identity, compact views, mutable state, and tasks remain separate

6. Non-goals and inappropriate uses

• Replacing field surveys or administrative microdata.
• Approximating or identifying real individuals.
• Simulating minors as public personas in v0.1.
• Treating synthetic narrative as literal biography truth.
• Making election or policy forecasts without external validation.

7. Construction pipeline

1. Household synthesis.
2. Adult assignment within households.
3. Geographic and life-stage allocation.
4. Education, labor, occupation, income, and socioeconomic assignment.
5. Migration and language-domain assignment.
6. Digital/media and cultural-profile assignment.
7. Civic, political, consumer, and latent-value assignment.
8. Weight calibration and disclosure tagging.
9. Persona-view rendering from structured fields.
10. Actor-state initialization.
11. Benchmark-task generation.

8. Schema overview

8.1 persona_core

• identity and linkage: synthetic_person_id, household_id, snapshot_id
• geography: region, municipality_class, urban_rural
• demography: age, age_group, gender, migration_background, years_in_spain_band
• education and work: education, labor_status, occupation_class, socioeconomic_tier, income
• language: first_language, language_of_identification, home_language, work_or_study_language, spanish_proficiency, immigrant_language_repertoire
• digital/media: device_access, internet_intensity, media_habit_cluster, primary_news_sources, platform_mix
• culture/leisure: reading_frequency, gaming_frequency, sports_affinity, cultural_interests, community_participation
• civic/political structure: turnout_propensity, institutional_trust, political_attention, ideology_interval_low, ideology_interval_high, issue_salience_top3
• consumer structure: decision_style, price_sensitivity, sustainability_orientation, local_purchase_preference, brand_loyalty
• governance: population_weight, benchmark_sampling_weight, uncertainty_level, disclosure_risk_level, field_provenance_ids

8.2 household_core

• composition: adult_count, minor_count, household_type, caregiving_role
• settlement: region, municipality_class, urban_rural
• economic context: tenure_band, housing_cost_burden, vehicle_access, consumption_constraint

8.3 persona_views

• micro_card
• standard_card
• policy_view
• consumer_view
• culture_view
• dialogue_view
• extended_profile for a stratified subset

8.4 actor_state_init

• current mood
• attention budget
• event sensitivity
• persuasion resistance
• memory style
• recent-media diet
• shopping context and preference volatility

8.5 benchmark_tasks

• task identity: id, family, split, synthetic_person_id
• binding: persona_view, scoring_target, allowed_context
• reproducibility: prompt_template_id, replay_seed
• task content: prompt

9. Provenance regime

Provenance tier	Meaning
Official statistics	Macro population anchors and structural baselines
Institutional or survey inputs	Conditioning inputs for language, media, civic, and social structure
Modeled latent variables	Values, motivations, and abstract behavioral tendencies
Rendered narrative	Spanish textual views derived from structured inputs

Companion provenance artifacts:

• source_registry.parquet for release-level source records
• field_provenance.parquet for field-group attribution

10. Quality controls

Control	Result
Region share MAE	0.022 pp
Age share MAE	2.95 pp
View budget compliance	100%
Benchmark families	9
Benchmark split types	4
High disclosure-risk rows	0.418%

11. Representative population structure

Signal	Result
First language = Spanish	76.896%
Migration background other than Spain-born	22.516%
Households with minors	38.013%
Private rent	39.471%
High housing-cost burden	29.676%
Tight consumption constraint	22.080%

12. Privacy and disclosure posture

Not published as stable public fields:

• exact birth dates
• street addresses
• email addresses
• phone numbers
• document numbers
• real employer names
• real school names
• stable public full-name columns

Narrative views are derived from structured fields and are designed for prompt usefulness rather than real-person imitation.

13. Current limitations

• The age profile remains the main demographic calibration gap.
• Household, media, and latent-value fields are synthetic abstractions and should be validated against the downstream task.
• The benchmark matrix is packaged, but live cross-model lift is not part of the release card.
• Disclosure flags help triage rows for review but are not a substitute for formal re-identification analysis.

14. Distribution and maintenance

• Data license: CC BY 4.0
• Release form: dated snapshot
• Default HF viewer config: persona_core
• Change policy: new versioned release rather than silent mutation of v0.1

15. Companion documents

• README.md
• EVALUATION_REPORT.md
• EVALUATION_METRICS.json
• PRIVACY_AND_DISCLOSURE.md

16. Citation

Cite the Hugging Face repository and the release id spain-reference-personas-2025-v0.1.