Add XL-SafetyBench datasets

[romanlutz]

Summary

Adds the XL-SafetyBench seed datasets to PyRIT. The benchmark contains 5,500 native-validated prompts across 10 country-language pairs and covers country-grounded jailbreak safety plus cultural sensitivity.

• Paper: arXiv:2605.05662
• Dataset: AIM-Intelligence/XL-SafetyBench on HuggingFace (CC-BY-4.0)
• Eval code: https://github.com/AIM-Intelligence/XL-SafetyBench

This PR is intentionally dataset-only. The benchmark-specific judges have been moved to a separate stacked follow-up.

What's in this PR

Three logical providers under pyrit/datasets/seed_datasets/remote/xl_safety_bench_dataset.py:

Provider	Emits	Size	Description
_XLSafetyBenchJailbreakDataset	SeedPrompt	4,500	Country-grounded adversarial attack prompts in five harm categories
_XLSafetyBenchJailbreakObjectivesDataset	SeedObjective	~1,500	Deduplicated harmful goals behind the attack prompts, for use with PyRIT attack strategies
_XLSafetyBenchCulturalDataset	SeedPrompt	1,000	Innocuous scenarios embedding local taboos or sensitivities in six categories

Each loader accepts countries: list[XLSafetyBenchCountry] | None; None loads all 10 countries, while callers can provide one or several enum values to select only those country files. This avoids separate per-country provider classes while retaining single-country and multi-country loading.

Shared enums cover the 10 countries, five jailbreak categories, six cultural categories, and language mode. Per-row provenance is preserved in metadata, including country/language display information, base queries, localized scenarios, and hidden violations.

Documentation

doc/code/datasets/1_loading_datasets cites XL-SafetyBench and its executed output lists the three logical benchmark providers.

Tests

tests/unit/datasets/test_xl_safety_bench_dataset.py covers all-country defaults, single- and multi-country filtering, country deduplication, objectives, category filtering, language selection, metadata propagation, schema validation, UTF-8 BOM normalization, and the three-name registry contract.

Notes for reviewers

• The upstream CSVs include a UTF-8 BOM on the id column; rows are normalized before field access so IDs and prompt names are retained.
• The jailbreak prompts loader has the default tag. Objectives and cultural loaders intentionally omit it to avoid duplicate/default sweep coverage.
• Cultural scenarios default to the local language and can also be loaded in English.
• CC-BY-4.0 attribution and the paper citation are preserved.
• The benchmark carries a content warning, consistent with similar safety datasets in PyRIT.

Deferred follow-ups

• XL-SafetyBench jailbreak and cultural judge scorers (separate stacked session)
• Benchmark scenarios
• Example notebook

Verification

• 79 focused dataset and registry metadata tests pass
• The loading-datasets notebook executes without errors, renders exactly the three logical XL-SafetyBench providers, and passes strict Jupytext round-trip validation
• Repository pre-commit hooks pass, including notebook sanitizers, Ruff, ty, and documentation checks

[hannahwestra25]

minor comment related to the title but lgtm otherwise!