sripad17/multilingual-safety-probe-test

Multilingual safety-probe test set (EN / RU / SR-Latn / SR-Cyrl / TE)

⚠️ Content warning. Half of the items are harmful prompts (violence and fraud requests, e.g. instructions for hurting people or scamming them), in five language versions. They are released for safety-evaluation research only. Reading them may be disturbing or upsetting.

Contents

1,000 rows = 200 curated English items × 5 language versions (English original + Russian, Serbian-Latin, Serbian-Cyrillic, Telugu translations). The 200 items are 100 harmful (50 violence / 50 fraud) and 100 benign (35 violence-adjacent near-miss, 35 fraud-adjacent near-miss, 30 clearly benign). Files: data/multilingual_test.parquet and an identical data/multilingual_test.csv.

column	type	meaning
item_id	string	stable id; prefix identifies the upstream dataset
language	string	en, ru, sr, te
script	string	en_Latn, ru_Cyrl, sr_Latn, sr_Cyrl, te_Telu
text	string	the prompt (final validator-corrected text for translations)
label	int	1 = harmful, 0 = benign
upstream_source	string	dataset the English item was selected from
upstream_license	string	license of that upstream dataset (see mix below)
validated	bool	row passed human review (see Validation; False for exactly 1 Telugu row)
correction_applied	bool	a human validator's correction replaced the machine translation (87 rows)

Provenance

upstream dataset	items (of 200)	rows (of 1,000)	license
nvidia/Aegis-AI-Content-Safety-Dataset-2.0	82 harmful	410	CC-BY-4.0
bench-llm/or-bench (or-bench-80k)	58 benign	290	CC-BY-4.0
databricks/databricks-dolly-15k	30 benign	150	CC-BY-SA-3.0
XSTest	12 benign	60	CC-BY-4.0
StrongREJECT	10 harmful	50	MIT
JailbreakBench/JBB-Behaviors	5 harmful	25	MIT
HarmBench	3 harmful	15	MIT

All 10 StrongREJECT items trace to MIT-licensed provenance within StrongREJECT (8 custom, 2 DAN); none come from its unlicensed upstream sources. English items were selected for unambiguous labels, translatability (no wordplay or culture-bound phrasing), and were decontaminated against WildGuardMix (the probe training source).

How translations were produced and validated

• Machine translation: claude-opus-4-8 (adaptive thinking), one language at a time, with LaBSE cross-lingual similarity gating: acceptance floor 0.75, up to 5 attempts per item; 100% of released rows scored ≥ 0.75.
• Serbian scripts: translated into Latin script, then transliterated to Cyrillic with cyrtranslit under a Latin→Cyrillic→Latin round-trip check; ambiguous nj/lj/dž digraphs were individually human-confirmed before any Cyrillic correction entered the dataset.
• Human validation: two native/proficient validators per language, each judging 150 items on three checks (meaning preserved / intensity preserved / reads naturally) with a 100-item overlap for agreement. Disagreements were resolved accept-or-correct with a deterministic tiebreak. Inter-annotator agreement (Gwet's AC1) on the overlap: Russian 0.785, Serbian 0.870, Telugu 0.855.
• Corrections: 87 of 800 translated rows carry a validator correction (correction_applied = True); text is always the final corrected version.
• Coverage: every translated row was human-reviewed except one Telugu row (aegis_f71a0126043c, left blank by its assigned validator), marked validated = False. English rows are the curated source items (validated = True refers to the curation review, not translation review).

Intended use

Evaluating cross-lingual generalization of safety classifiers and activation-space harm probes — in particular recall gaps between English and lower-resource languages/scripts, and threshold-calibration effects. Not a training set; not a jailbreak benchmark (all prompts are vanilla, direct requests).

Licensing and attribution

The whole set is released under CC-BY-SA-4.0 (https://creativecommons.org/licenses/by-sa/4.0/). Share-alike is required by the 30 Dolly-derived items (CC-BY-SA-3.0, upgraded per its §4(b) compatibility path) and is compatible with the CC-BY-4.0 and MIT upstream sources. The upstream_license column preserves the per-item mix; MIT copyright notices for StrongREJECT, JailbreakBench, and HarmBench are in NOTICE. Translations are adaptations of the upstream English prompts; modifications (translation + validator corrections) are indicated per row.

Attribution: NVIDIA (Aegis 2.0, itself derived from Anthropic hh-rlhf, MIT), OR-Bench (Cui et al.), Databricks (dolly-15k; Wikipedia-derived passages also credit Wikipedia contributors, CC-BY-SA), XSTest (Röttger et al.), StrongREJECT (Souly et al.), JailbreakBench (Chao et al.), HarmBench (Mazeika et al.).