FEAT: Add sibling loaders for dataset variants pulling distinct upstream artifacts

[romanlutz]

Why

tests/end_to_end/test_all_datasets.py parametrizes over every registered SeedDatasetProvider but only ever calls provider_cls() with no arguments. Loaders whose non-default constructor arguments select between distinct upstream artifacts — a different HF split, a different JSONL URL, a different instr-resp field on the same row, a different HF config — therefore only have their default config exercised against real upstream data. A schema drift in a non-default variant (CSV column rename, missing split, image directory rename) goes unnoticed until a user hits it in production. Unit tests catch the wiring; they do not catch upstream drift.

What

Apply the OR-Bench-style aliasing pattern across the 8 affected loaders on main: each distinct upstream variant becomes its own registered SeedDatasetProvider subclass with the relevant constructor default pinned and a unique dataset_name. The registry-driven e2e parametrization then picks up every new subclass automatically — no test-runner refactor needed.

Per-loader breakdown (25 new registered subclasses):

Loader	New siblings	Distinct upstream selector
_AyaRedteamingDataset	+7 (Hindi, French, Spanish, Arabic, Russian, Serbian, Tagalog)	per-language JSONL URL
_BabelscapeAlertDataset	+1 (_BabelscapeAlertOriginalDataset)	HF config (alert vs default alert_adversarial)
_CategoricalHarmfulQADataset	+2 (Chinese, Vietnamese)	HF split
_HiXSTestDataset (gated)	+1 (English)	column read (Hindi vs English prompt)
_MSTSDataset	+10 (German, Russian, Chinese, Hindi, Spanish, Italian, French, Korean, Arabic, Farsi)	HF split
_SaladBenchDataset	+2 (attackEnhanced, defenseEnhanced)	HF split
_VLGuardDataset (gated)	+2 (SAFE_UNSAFES, SAFE_SAFES)	instr-resp field read on the same metadata zip

Loaders left unchanged because the default already fans out across every upstream artifact: EquityMedQA (subset_name="all" iterates 11 HF configs), MedSafetyBench (same pattern), ComicJailbreak (default fetches all 5 templates), VLSU / HarmBenchMultimodal / VisualLeakBench (all params are post-fetch filters), etc.

Notes

• Backward compatible at the class level. Existing kwarg-based selection on parent loaders is untouched. No deprecations on the Python API.
• E2E CI cost. The image-heavy MSTS (10 new variants) and VLGuard (2 new variants) additions could blow up the daily e2e download volume. To bound this, tests/end_to_end/test_all_datasets.py grew a small _CAPPED_PROVIDERS set that applies max_examples=6 in e2e only. Production defaults are unchanged.
• Aya legibility. The original _AyaRedteamingDataset inlined a giant Literal[...] for harm_categories. Restating that union eight times for each subclass would have been unreadable, so it was factored into module-level AyaLanguage / AyaHarmCategory type aliases that both the parent and all 7 subclasses use.
• Unit-test coverage. Each new subclass has smoke coverage in the matching tests/unit/datasets/ file: dataset_name, correct pinned config/split/subset forwarded to the parent fetch method, kwarg forwarding for max_examples and token.
• Docs notebook. doc/code/datasets/1_loading_datasets.ipynb's rendered dataset-name list cell was surgically updated in place so the docs site picks up the new variants without a full re-execution introducing environmental noise (timestamps, UUIDs, Windows paths) into the diff.
• Gated loaders. _HiXSTest* and _VLGuard* variants inherit the parent's HUGGINGFACE_TOKEN plumbing and will skip in CI without the secret, same behavior as today.

Out of scope

FigStep is on a separate branch (romanlutz/figstep-dataset-loader) and is not yet on main. It is the motivating example for this audit but the refactor is handled in that PR, where the same pattern can be applied more cleanly (clean OR-Bench-style abstract base + 3 concrete subclasses with no deprecation surface).

Naming follow-up

After adding sibling subclasses, the parent classes' generic dataset_names (aya_redteaming, salad_bench, babelscape_alert, categorical_harmful_qa, hixstest, vlguard) sat awkwardly next to their newly-added language/split-specific siblings — there was no way to tell from the registry which variant the generic name referred to without reading source. Each parent's dataset_name is now renamed to spell out the variant it actually fetches by default:

Parent (was)	Parent (now)	Pinned default
aya_redteaming	aya_redteaming_english	language="English"
salad_bench	salad_bench_base	split="base"
babelscape_alert	babelscape_alert_adversarial	category="alert_adversarial"
categorical_harmful_qa	categorical_harmful_qa_english	language="en"
hixstest	hixstest_hindi	reads Hindi column
vlguard	vlguard_unsafes	subset=UNSAFES

MSTS is intentionally not renamed: its parent's default (size="large") is the all-11-languages aggregate, semantically distinct from any per-language sibling, so msts as the aggregate name remains correct.

Note: This changes the public dataset_name strings for these six loaders. PyRIT is pre-1.0; any external code calling SeedDatasetProvider.get_by_name("aya_redteaming") (etc.) will need to switch to the new explicit name.

Verification

• uv run pytest tests/unit/datasets/ -q — 576 tests pass
• uv run ruff check pyrit/ tests/ — All checks passed
• uv run ruff format --check pyrit/ tests/ — All files already formatted
• Registry sanity check: SeedDatasetProvider.get_all_providers() returns 91 providers (was 66), all 25 new subclasses register with unique dataset_names.