{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.01322", "kind": "paper", "url": "https://arxiv.org/abs/2606.01322"}], "authors": ["Victor Akinode", "Senyu Li", "Wassim Hamidouche", "Waqas Zamir", "Inbal Becker-Reshef", "David Ifeoluwa Adelani"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.01322", "source_url": "https://arxiv.org/abs/2606.01322"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "TUKABENCH is a benchmark for jailbreak evaluation in seven African languages, extending JailbreakBench with human translations, cultural adaptations, and code-switched prompts. It assesses LLM safety in low-resource languages using metrics like Refused, Jailbroken, and Deflection, with human validation of LLM-as-a-judge."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.01322", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.01322", "title": "TukaBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.01322", "modality": null, "name": "TukaBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_tukabench_beb4bae4", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "54243a37989beac4e91ff83dc837ce15328e71b3e35e74bdfba9a5ddd949d85a", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.01322"}, "publisher": null, "released": "2026-05-31", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.01322", "source_url": "https://arxiv.org/abs/2606.01322"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-01322", "source": "claire_radar", "source_benchmark_id": "2606.01322", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 0, "hfPaperUrl": "https://huggingface.co/papers/2606.01322", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-09-30", "state": "stale"}}, "status": "stale"}, "capabilities": [], "confidence": "High", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "10d2859030b484f55598584f77c60bfaeaac1940867c15577f58e92d3b7d502c", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "TUKABENCH is a benchmark for jailbreak evaluation in seven African languages, extending JailbreakBench with human translations, cultural adaptations, and code-switched prompts. It assesses LLM safety in low-resource languages using metrics like Refused, Jailbroken, and Deflection, with human validation of LLM-as-a-judge.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce TUKABENCH, a jailbreak benchmark for seven African languages that extends JailbreakBench (JBB) beyond direct translation through four settings: human translation of JBB prompts, English adaptation to African contexts followed by human translation, human-curated prompts validated through interactions with GPT-5.2, and code-switched prompts combining English and African languages, isolating the effect of language, cultural grounding, and prompt evasiveness on model safety."}, "familyId": "bmf_65c1dc6edc86", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_tukabench_beb4bae4", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.01322", "paper": "https://arxiv.org/abs/2606.01322", "pdf": "https://arxiv.org/pdf/2606.01322", "project": null, "report": "https://arxiv.org/abs/2606.01322"}, "metrics": [], "motivation": "Safety evaluation of Large Language Models (LLMs) remains heavily English-centric, leaving Low-Resource Languages (LRLs), particularly African ones, critically underexplored.", "name": "TukaBench", "oneLine": "TUKABENCH is a benchmark for jailbreak evaluation in seven African languages, extending JailbreakBench with human translations, cultural adaptations, and code-switched prompts. It assesses LLM safety in low-resource languages using metrics like Refused, Jailbroken, and Deflection, with human validation of LLM-as-a-judge.", "paperTitle": "TukaBench: A Culturally Grounded Jailbreak Benchmark for African Languages", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.01322", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-31", "source": {"authors": ["Victor Akinode", "Senyu Li", "Wassim Hamidouche", "Waqas Zamir", "Inbal Becker-Reshef", "David Ifeoluwa Adelani"], "categories": ["cs.CL", "cs.AI"], "id": "2606.01322", "title": "TukaBench: A Culturally Grounded Jailbreak Benchmark for African Languages", "type": "arxiv", "url": "https://arxiv.org/abs/2606.01322"}, "sourceUpdatedAt": "2026-05-31T16:14:37Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "The benchmark addresses a gap in safety evaluation for low-resource African languages, showing that models are more vulnerable to jailbreak prompts in these languages. It could inform safer deployment of LLMs in multilingual contexts."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
