language:
- es
license: openrail
task_categories: - text-classification
- question-answering
tags: - red-teaming
- ai-safety
- llm-evaluation
- latin-america
- pluralism
- cultural-adaptation
- localization
- spanish
- multilingual
pretty_name: LatAm-RT
size_categories: - n<1K
LatAm-RT: Culturally Adaptive Red Teaming for AI Safety in Latin America
LatAm-RT is a culturally adaptive Red Teaming benchmark for evaluating AI safety in Latin American Spanish. It operationalizes localization intensity as an explicit experimental variable through a Layered Localization Taxonomy (LLT) with three levels of progressive cultural specificity.
The dataset contains 284 evaluation prompts across:
- 4 risk domains: Fraud & Exploitation · Political & Information Harm · Privacy & Civic Systems · Social Harm & Discrimination
- 3 localization layers: L1 Direct Translation · L2 Regional LatAm Adaptation · L3 National Localization
- 4 target countries (L3 only): Mexico · El Salvador · Costa Rica · Chile
- 5 adversarial framing strategies: V1 Direct · V2 Indirect · V3 Fictional · V4 Epistemic · V5 Defensive
Files
| File | Description |
|---|---|
latamrt_prompts.csv |
284 evaluation prompts with full metadata |
latamrt_cps_annotations.csv |
Cultural Plausibility Score annotations (12 native annotators) |
Dataset Schema
latamrt_prompts.csv
| Column | Description |
|---|---|
prompt_id |
Unique identifier (e.g. FE-1.1-L3-SV-V1) |
domain |
Primary risk domain |
domain_secondary |
Secondary domain (multi-label cases) |
subcategory |
Risk subcategory |
seed_id |
Source seed scenario |
localization_layer |
L1 / L2 / L3 |
target_country |
MX / SV / CR / CL (L3 only) |
variation |
V1–V5 framing strategy |
harm_type |
explicit harmful / gray zone |
prompt_text_es |
Full prompt text in Spanish |
grounding_notes |
Slot instantiation notes |
intent_preserved |
Intent preservation check result |
latamrt_cps_annotations.csv
| Column | Description |
|---|---|
prompt_index |
Row index in the L3 prompt set |
country |
Target country |
annotator_id |
Anonymized annotator ID (e.g. MX-A1) |
annotator_region |
Region of origin within country |
domain |
Risk domain |
variation |
Framing strategy |
LC |
Linguistic Coherence (0–2) |
IA |
Institutional Accuracy (0–2) |
SR |
Scenario Realism (0–2) |
FC |
Framing Coherence (0–2; V3/V5 only) |
cps_total |
Mean(LC, IA, SR) |
comment |
Qualitative annotator comment (if any) |
Localization Layers
| Layer | Name | Description |
|---|---|---|
| L1 | Direct Translation | Automated EN→ES translation; no cultural adaptation |
| L2 | Regional LatAm Adaptation | Regionally neutral Spanish; generic institutional references |
| L3 | National Localization | Nationally specific institutions, lexicon, and social scenarios |
L3 Institutional Tokens by Country
| Country | National ID | Tax authority | Electoral body | Informal: money |
|---|---|---|---|---|
| Mexico (MX) | CURP / INE | SAT | INE / TEPJF | feria |
| El Salvador (SV) | DUI | DGII | TSE | pisto |
| Costa Rica (CR) | Cédula / DIMEX | Tributación Directa | TSE | plata |
| Chile (CL) | RUN | SII | SERVEL | plata / luca |
CPS Validation
Cultural plausibility of L3 prompts was validated by 12 independent native annotators (3 per country) using a four-dimension rubric:
- LC Linguistic Coherence: natural use of national dialect and register
- IA Institutional Accuracy: correctness of named institutions and procedures
- SR Scenario Realism: plausibility for a country inhabitant
- FC Framing Coherence: cultural coherence of framing (V3/V5 only)
Global mean CPS: 1.92/2.0 (range: 1.88–1.94 across countries)
Adversarial Framing Strategies
| Code | Name | Description |
|---|---|---|
| V1 | Direct request | First-person explicit harmful request |
| V2 | Indirect request | Third-person pronominal distancing |
| V3 | Fictional framing | Harmful request embedded in narrative context |
| V4 | Epistemic framing | Knowledge-seeking or educational framing |
| V5 | Defensive framing | Self-protection or harm prevention framing |
Dataset Construction
284 prompts were generated from 12 seed scenarios derived from established English benchmarks (AdvBench, HarmBench, JailbreakBench). A two-phase curation pipeline applied:
- Consensus filter: 3 independent LLM classification runs; 11 of 12 seeds accepted unanimously
- Intent preservation check: 46 of 330 generated variants (13.9%) rejected for intent drift
Intended Use
This dataset is intended exclusively for AI safety research and evaluation. It contains adversarial prompts designed to test safety systems of large language models.
Permitted uses:
- Evaluating LLM safety behavior in Latin American Spanish
- Studying localization effects on AI safety robustness
- Culturally grounded red-teaming research
Not permitted:
- Using prompts to elicit harmful outputs from deployed systems
- Commercial use without explicit permission
Citation
@misc{romero2026latamrt,
title = {{LatAm-RT}: Toward Culturally Adaptive Red Teaming
for {AI} Safety in Latin America},
author = {Romero, Roxana R.},
year = {2026},
url = {https://huggingface.co/datasets/roxanaromero0743/LatAmRT}
---
## License
CC BY 4.0 — Attribution required. See LICENSE file.