apol/spain-reference-personas-frontier

Spain Reference Personas Frontier

Spain Reference Personas Frontier is a synthetic reference population and benchmark substrate for Spanish-language LLM work grounded in the territorial, household, cultural, linguistic, and civic structure of Spain. It is built for simulation, evaluation, prompt conditioning, and scenario analysis.

The release is inspired by NVIDIA's Nemotron Personas line, especially the multi-view packaging visible in Nemotron Personas USA, while extending that idea into a benchmark-oriented reference package for Spain.

Release snapshot

Item	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
LLM-facing views	6,350,524
Actor-state rows	1,000,000
Benchmark tasks	1,800
Extended-profile coverage	35.1%
Total package rows	8,889,089
Total package size	5.584 GB

Why this package exists

• Controllability: structured fields support filtering, weighting, and subgroup analysis directly.
• Behavioral usefulness: the release includes actor state and benchmark tasks rather than only descriptive prose.
• Token efficiency: every public view has a declared budget and measured compliance.
• Reproducibility: release metadata, task splits, replay seeds, and evaluation summaries are explicit.
• Household realism: adults remain linked to tenure, burden, caregiving, and consumption context.

Use cases

Audience	Example workflow	Start with
Sociologists	Slice the population by region, language, household form, migration background, and values before designing fieldwork or interview sampling	persona_core, household_core
Poll analysts	Retrieve a cohort, attach policy_view, and compare open-ended synthetic answers across held-out splits	persona_core, persona_views, benchmark_tasks
Policy analysts	Simulate reactions to housing, care, labor, migration, or cost-of-living interventions using stable structure plus mutable state	household_core, actor_state_init, persona_views
Economists	Study household burden, consumption constraints, price sensitivity, and tenure differences in consumer-choice prompts	household_core, persona_core, consumer_view
Media researchers	Model trust, platform exposure, recent-media pathways, and event-response heterogeneity	persona_core, actor_state_init, dialogue_view
Model builders	Benchmark compact versus extended views with explicit held-out persona and held-out task regimes	all configs

Example programs

• Housing policy reaction studies split by tenure, burden, age, and region.
• Synthetic polling stress-tests that compare short survey answers against richer policy views.
• Consumer trade-down simulations under inflation using price sensitivity and household constraints.
• Regional culture and language robustness tests for Spanish-first models serving co-official-language contexts.
• Event-reaction experiments where the same stable persona receives different recent-media states.
• Multi-turn family, workplace, or community interaction tasks that require stable persona identity plus mutable memory.

What ships

Config	Rows	Role
persona_core	1,000,000	Stable adult structure, weights, language profile, civic profile, consumer profile, value axes, provenance ids
household_core	536,741	Household composition, tenure, burden, vehicle access, caregiving, consumption constraints
persona_views	6,350,524	micro_card, standard_card, policy_view, consumer_view, culture_view, dialogue_view, optional extended_profile
actor_state_init	1,000,000	Mood, attention, event sensitivity, persuasion resistance, memory style, recent-media diet
benchmark_tasks	1,800	Task prompts, scoring targets, replay seeds, split metadata, recommended persona view
source_registry	11	Release-level source inventory
field_provenance	13	Field-group provenance map

Package logic

1. Filter or weight cohorts in persona_core.
2. Join household_core when housing or economic context matters.
3. Attach the smallest useful persona view from persona_views.
4. Add actor_state_init only when recency, mood, or event exposure matter.
5. Score behavior with benchmark_tasks instead of relying on anecdotal prompt outputs.

Evaluation at a glance

Metric	Result	Interpretation
Region share MAE	0.022 pp	Tight regional alignment for macro subgroup work
Region max absolute error	0.051 pp	No large regional drift in the released person table
Age share MAE	2.95 pp	Main remaining calibration gap in v0.1
Age max absolute error	4.16 pp	Largest deviation is in middle-age representation
View budget compliance	100%	All public views stay inside their declared limits
Benchmark matrix	9 families / 4 splits	Explicit generalization structure exists in the bundle
Weight spread	0.9889 - 1.0551	Weights remain mild instead of extreme
High disclosure risk	0.418%	Small review tail remains visible as metadata

Observed regional shares

Andalucía              17.90%  ##################
Cataluña               16.35%  #################-
Madrid                 14.19%  ##############----
Com. Valenciana        10.62%  ###########-------
Galicia                 5.73%  ######------------
Castilla y León         5.07%  #####-------------
País Vasco              4.70%  #####-------------
Canarias                4.58%  #####-------------

Age calibration summary

Age group	Target	Observed	Error
18-24	8.0%	10.484%	+2.48 pp
25-34	13.0%	16.996%	+4.00 pp
35-44	17.0%	15.843%	-1.16 pp
45-54	19.0%	14.840%	-4.16 pp
55-64	17.0%	13.460%	-3.54 pp
65+	26.0%	28.377%	+2.38 pp

View-layer efficiency

View	Count	Avg tokens	Max tokens	Utilization	Pass rate
micro_card	1,000,000	99.8	120	83.1%	100.0%
standard_card	1,000,000	175.7	212	70.3%	100.0%
policy_view	1,000,000	89.2	97	49.5%	100.0%
consumer_view	1,000,000	95.6	113	53.1%	100.0%
culture_view	1,000,000	105.8	153	58.8%	100.0%
dialogue_view	1,000,000	83.4	93	46.3%	100.0%
extended_profile	350,524	364.5	407	60.8%	100.0%

micro_card         99.8 / 120   ############--
standard_card     175.7 / 250   ##########----
policy_view        89.2 / 180   #######-------
consumer_view      95.6 / 180   #######-------
culture_view      105.8 / 180   ########------
dialogue_view      83.4 / 180   ######--------
extended_profile  364.5 / 600   ########------

Household and economic context

Signal	Result
Average adults per household	1.863
Average minors per household	0.560
Households with minors	38.013%
Private rent	39.471%
Mortgage	21.837%
Owner outright	21.602%
High housing-cost burden	29.676%
Tight consumption constraint	22.080%

Housing-cost burden	Share
moderate	36.710%
low	33.614%
high	29.676%

Benchmark design

Benchmark family	Tasks
policy_opinion	200
election_turnout	200
poll_response	200
event_reaction	200
media_trust	200
consumer_choice	200
culture_identity	200
multi_turn_social	200
future_expectations	200

Split regime	Tasks
in_distribution	450
heldout_persona_seen_task	450
seen_persona_heldout_task	450
heldout_persona_heldout_task	450

Loading

from datasets import load_dataset

personas = load_dataset(
    "apol/spain-reference-personas-frontier",
    "persona_core",
    split="train",
    token=True,
)

Limits and cautions

• This is a synthetic reference population, not observed microdata.
• The package is suitable for simulation and evaluation, not for replacing field surveys.
• Age calibration remains the main statistical weakness of v0.1.
• High-disclosure-tagged rows are exposed as metadata so downstream users can exclude them when needed.
• Live cross-model benchmark lift is not claimed in the card itself; the bundle ships the infrastructure needed to run it reproducibly.

Companion documents

• DATASHEET.md
• EVALUATION_REPORT.md
• PRIVACY_AND_DISCLOSURE.md
• EVALUATION_METRICS.json

Citation

If you use this dataset, cite the repository and the release id spain-reference-personas-2025-v0.1.