FEAT: Backfill class-level metadata for all remote seed datasets

[romanlutz]

What

Adds class-level tags, size, and modalities to every remote seed dataset loader so they participate in SeedDatasetFilter discovery (e.g. SeedDatasetFilter(tags={"default"})). Before this change, only 5 of ~33 remote loaders declared metadata, so the others were silently skipped by metadata-driven filtering.

This is the follow-up to the review discussion on #1757 (cc @jsong468) where reviewer asked why some loaders declared these fields and others didn't. Answer: most predate the metadata schema and just hadn't been backfilled.

How

1. Pinned a canonical, advisory tag vocabulary as RECOMMENDED_TAGS in pyrit/datasets/seed_datasets/seed_metadata.py. Users can still set custom tags — the metadata parser does not enforce, but a new parametrized coverage test does.
2. Documented the 5-condition rule for the special default tag inline in seed_metadata.py:
   1. Ungated — no HF token, API key, auth, or signup.
   2. Citable — peer-reviewed paper / established benchmark.
   3. Single-call — await loader.fetch_dataset_async() works with no manual setup.
   4. Size >= medium (>=100 prompts).
   5. Broadly applicable — not narrowly scoped to a vertical (medical, legal, cybersecurity). Cross-cutting axes like privacy, bias, multimodal, multilingual, refusal, and jailbreak DO count.
3. Walked every remote loader and assigned size / tags / modalities based on the loader's docstring, tests, and upstream dataset card. Added inline # N prompts comments next to each size so reviewers can verify the bucket choice locally.
4. Renamed the non-canonical multilingual_culture tag on _SGXSTestDataset to multilingual and dropped its default tag (SGXSTest is gated on HF).
5. Marked _ORBenchBaseDataset as should_register = False (it has no usable dataset_name) and explicitly opted the three OR-Bench leaf classes back in.
6. Added TestRemoteLoaderMetadataCoverage — a parametrized test that walks every concrete _RemoteDatasetLoader subclass via auto-registration and asserts: metadata is present, tags/size/modalities are non-empty, size is in SeedDatasetSizeCategory, and tags is a subset of RECOMMENDED_TAGS (catches future typos like multilingual_culture).

Class-level harm_categories is intentionally deferred — per-row SeedPrompt.harm_categories already labels individual prompts; picking a "broadest" class-level summary is a judgment call better made by domain owners in a focused follow-up.

Backfill table

Loader	size	modalities	tags	default?
_AegisContentSafetyDataset	huge	(text,)	{default, safety}	YES
_AyaRedteamingDataset	medium	(text,)	{safety, multilingual}	no — per-language ~few hundred
_BabelscapeAlertDataset	huge	(text,)	{default, safety, jailbreak}	YES
_BeaverTailsDataset	huge	(text,)	{default, safety}	YES
_CBTBenchDataset	medium	(text,)	{safety, medical}	no — vertical
_CCPSensitivePromptsDataset	small	(text,)	{safety, multilingual}	no
_DarkBenchDataset	medium	(text,)	{default, safety}	YES
_EquityMedQADataset	medium	(text,)	{safety, bias, medical}	no — vertical
_ForbiddenQuestionsDataset	medium	(text,)	{default, safety, jailbreak}	YES
_HarmBenchMultimodalDataset	small	(text, image)	{safety, jailbreak, multimodal}	no
_HarmfulQADataset	large	(text,)	{default, safety, jailbreak}	YES
_JBBBehaviorsDataset	small	(text,)	{safety, jailbreak}	no
_LibrAIDoNotAnswerDataset	medium	(text,)	{default, safety, refusal}	YES
_LLMLatentAdversarialTrainingDataset	large	(text,)	{default, safety, jailbreak}	YES
_MedSafetyBenchDataset	large	(text,)	{safety, medical}	no — vertical
_MLCommonsAILuminateDataset	large	(text,)	{default, safety}	YES
_MultilingualVulnerabilityDataset	medium	(text,)	{default, safety, multilingual}	YES
_ORBench80KDataset	huge	(text,)	{default, safety, refusal}	YES
_ORBenchHardDataset	large	(text,)	{default, safety, refusal}	YES
_ORBenchToxicDataset	large	(text,)	{default, safety, refusal}	YES
_PKUSafeRLHFDataset	huge	(text,)	{default, safety}	YES
_PromptIntelDataset	medium	(text,)	{safety, jailbreak, cybersecurity}	no — API key
_RedTeamSocialBiasDataset	small	(text,)	{safety, bias, multiturn}	no
_SaladBenchDataset	huge	(text,)	{default, safety, jailbreak}	YES
_SimpleSafetyTestsDataset	small	(text,)	{safety}	no
_SorryBenchDataset	large	(text,)	{safety, jailbreak, synthetic}	no — gated
_SOSBenchDataset	large	(text,)	{safety, medical, cybersecurity}	no — vertical
_TDC23RedteamingDataset	small	(text,)	{safety, jailbreak}	no
_ToxicChatDataset	huge	(text,)	{default, safety, multiturn}	YES
_TransphobiaAwarenessDataset	medium	(text,)	{default, safety, bias}	YES
_VLGuardDataset	large	(text, image)	{safety, multimodal}	no — gated
_VLSUMultimodalDataset	large	(text, image)	{default, safety, multimodal}	YES
_XSTestDataset	medium	(text,)	{default, safety, refusal}	YES
_SGXSTestDataset (fixed)	medium	(text,)	{safety, multilingual}	no — gated, was {default, safety, multilingual_culture}

Already-tagged (unchanged): _HarmBenchDataset, _ComicJailbreakDataset, _VisualLeakBenchDataset.

No-op

• No public API changes
• No runtime behavior changes
• No changes to per-row SeedPrompt.harm_categories
• The 4 already-tagged loaders are not migrated to immutable frozenset / tuple style (cosmetic — out of scope)

Discussion link

#1757

[behnam-o]

couple of minor comments, but also looks good as is.