Dataset name stringlengths 11 180 | Year stringlengths 4 108 | Language(s) covered stringlengths 11 272 | Resource level stringclasses 7
values | Source paper (arXiv/DOI) stringlengths 18 135 | Dataset link (HF/GitHub) stringlengths 34 294 | Authors stringlengths 11 556 | Model Evaluated stringlengths 5 514 | Format / structure stringlengths 10 327 | Method of creation stringlengths 9 363 | Translation direction stringlengths 9 300 | Harm categories covered stringlengths 24 725 | Harm category gaps stringlengths 21 345 | Global vs local harm stringlengths 2 332 | Who annotated it stringlengths 6 323 | Native review of THIS language? stringlengths 2 302 | Dialect / cultural handling stringlengths 10 264 | IAA score (this slice) stringlengths 17 366 | Size for THIS language stringlengths 28 325 | Rationale for the call stringlengths 70 407 | Quality concerns / oddities stringlengths 55 1.14k | Include / Exclude / Maybe + flag stringlengths 7 262 | License / usage rights stringlengths 9 223 | Access status stringlengths 6 204 | Domain / source stringlengths 9 239 | Script stringclasses 7
values | Code-switching coverage stringlengths 2 237 | Native taxonomy (verbatim) stringlengths 48 483 | Annotator count & background stringlengths 49 353 | Annotation guidelines public stringlengths 7 239 | Reported per-language performance stringlengths 56 564 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
[FR] Aya Redteaming | 2024 | French (+7) | High | arxiv.org/abs/2406.18682 | https://huggingface.co/datasets/CohereLabs/aya_redteaming | Cohere Labs | Aya 23 (8B), Aya 23 (35B), GPT-4 | prompt-only | Native-authored | n/a (native) | hate, harassment, self-harm, sexual, violence | (broad — few FR gaps) | Yes | Native speakers | Yes — native review | Monolithic | reported per-lang | ~900 FR prompts (this slice) | FR is in Aya's native core — written directly by French native speakers, not translated. | Rare case where collection headline = language reality. | Include | Apache 2.0 (HF dataset card) | Public (Hugging Face) | Human-authored adversarial prompts (paid native speakers per language) | Latin | No (monolingual French prompts) | bullying & harassment, discrimination & injustice, graphic material, hate speech, non-consensual sexual content, self-harm, violence — each flagged global vs local harm | Paid French native speakers (compensated annotators; per-language teams) | Partial (methodology documented in paper; full guidelines not released) | Yes — per-language prompt counts (~900 FR) + global-vs-local harm distributions reported per language |
[HA] AfriHate (collection-level row) | 2025 | Hausa, Swahili (+13) //// Hausa (hau), Swahili (swa), Algerian Arabic (arq), Amharic (amh), Igbo (ibo), Kinyarwanda (kin), Moroccan Arabic (ary), Nigerian Pidgin (pcm), Oromo (orm), Somali (som), Tigrinya (tir), Twi (twi), isiXhosa (xho), Yorùbá (yor), and isiZulu (zul) | Low | https://arxiv.org/abs/2501.08284 | https://huggingface.co/datasets/afrihate/afrihate | Shamsuddeen Hassan Muhammad, Idris Abdulmumin, Abinew Ali Ayele, David Ifeoluwa Adelani, Ibrahim Said Ahmad, Saminu Mohammad Aliyu, Paul Röttger, Abigail Oppong, Andiswa Bukula, Chiamaka Ijeoma Chukwuneke, Ebrahim Chekol Jibril, Elyas Abdi Ismail, Esubalew Alemneh, Hagos Tesfahun Gebremichael, Lukman Jibril Aliyu, Meri... | AfroXLMR (continued pretraining + LID adaptation), Aya-23-8B, mBERT, XLM-R, AfriBERTa variants and other multilingual baselines used in benchmark experiments. | Social media posts annotated with three primary labels (Hate, Abusive, Neutral). Hate instances receive an additional target annotation (ethnicity, religion, gender, politics, disability, other). | expert/native annotation | None. Data are native-language social media posts, not translations from English. | Hate Speech; Abusive Language. Hate targets: Disability, Ethnicity, Gender, Politics, Religion, Other. | Does not explicitly cover misinformation, self-harm, suicide, cybercrime, scams, privacy violations, child safety, sexual exploitation, jailbreak prompts, biosecurity, or violence outside hate/abuse. | Strong combination. Covers universal hate speech while preserving locally meaningful targets, political identities, ethnic groups, and culturally specific abusive language. | Native speakers | Yes — Native speakers familiar with the language and culture performed annotation | Strong emphasis on cultural competence. The paper explicitly discusses code-mixing, digraphia, regional linguistic variation, and culturally dependent hate expressions. However, dialect-level analyses are limited. | Free-marginal Randolph κ (Table 4): Hausa 0.75, Swahili 0.55 (collection range 0.46–0.81) | Hausa ≈6,644 posts; Igbo ≈5,003; Yoruba ≈3,700; Swahili >21,000. | Designed to provide the largest unified multilingual African hate speech benchmark, overcoming fragmentation of previous datasets and enabling cross-language evaluation. | Dataset sizes are highly imbalanced; annotation agreement varies considerably (0.46–0.81); data originate from multiple independent collection protocols; some datasets were aggregated while others were newly collected; hate labels remain relatively sparse in several languages. | Include | Apache 2.0 (HF card); gated access agreement on HF | Public Available but GATED on HF (login + accept conditions to access files); GitHub repo open | Twitter/X | Latin (majority); Ethiopic (Amharic, Tigrinya); Arabic script (Algerian/Moroccan Arabic) | Some — Nigerian Pidgin included; code-mixing noted in collection; per-slice varies Present. Authors explicitly discuss code-mixing as a characteristic challenge of African social media. | Hate, Abusive, Neutral. Hate targets: Disability, Ethnicity, Gender, Politics, Religion, Others. | Per Table 4 — Hausa: 3 annotators total, 3/instance (manual collection + pre-annotation); Swahili: 5 total, 3/instance; native speakers, majority vote | Yes (annotation scheme + per-language lexicons released). Annotation guidelines are described in the paper and appendix. | Yes — per-language baselines (Afri-centric PLMs + prompted LLMs); free-marginal multirater κ per language (Table 4) |
[SW] RTP-LX | 2024 (Swahili added V0.3, Jan 2024) | Swahili + 37 others (38 total; SW has the FULL component set incl. Completions & culturally-specific prompts — not in the reduced asterisk tier) | Mid | https://arxiv.org/abs/2404.14397 | https://github.com/microsoft/RTP-LX | de Wynter et al. (Microsoft; AAAI 2025, pp. 27940–27950) | GPT-4 | prompt-only (+ synthetic completions, human-annotated; benign completions) | Transcreated (native speakers) | EN→Swahili (transcreated, not literal MT) | Bias, Insult, Identity Attack, Microaggression, Violence, Self-Harm, Sexual Content, Overall Toxicity (8 dims) | misinformation/extremism not explicit dims; harassment folded into insult/identity-attack | Yes (culturally-specific prompt subset included for Swahili) | Professional transcreators + human annotators (majority vote) | Yes — native review | Translators encouraged to use dialects & mark them in the corpus | Cohen's κw per language; low model–human agreement on subtle dims is a headline finding | ~1,100 Swahili prompts (1k+ per locale) + completions | Include — Swahili shipped in V0.3 (Jan 2024) with the full component set; the transcreated side of the Swahili provenance contrast vs native AfriHate. | Language verification: RTP-LX's repository lists 38 covered languages; Hausa is not among them, and Swahili was added in release V0.3 (January 2024) with the full component set. This entry therefore records the Swahili slice. | Include | Research-only (Microsoft; entries zipped & password-protected) | Public (GitHub, password-gated zips) | Reddit-seeded RealToxicityPrompts, professionally transcreated into Swahili | Latin | No (monolingual Swahili; dialect indications allowed) | Bias, Insult, Identity Attack, Microagression, Violence, Self-harm, Sexual content, Overall toxicity | Professional transcreators + human annotators, majority vote; multi-national team | Partial (harm-category definitions in paper; full guidelines partial) | Yes — per-language Cohen's κw reported; SW in per-language results |
[HA] A MULTILINGUAL DATASET FOR OFFENSIVE
LANGUAGE AND HATE SPEECH DETECTION FOR
HAUSA, YORUBA AND IGBO LANGUAGES | 2024 | Hausa (+2) //// Igbo, Yoruba | Low | https://arxiv.org/abs/2406.02169 | None available. Authors state they plan to release the dataset and models publicly. No repository is provided | Saminu Mohammad Aliyu, Gregory Maksha Wajiga, Muhammad Murtala | XLM-RoBERTa-base, BERT-base-Multilingual-Cased, Morit/XLM-T-Full-XNLI, Davlan/Naija-Twitter-Sentiment-Afriberta-Large | prompt-only | Native tweets collected through keyword-based Twitter Academic API search, followed by manual annotation by native speakers. | None. Native-authored tweets only | Hate, Offensive Language | Does not cover harassment as a separate category, self-harm, sexual content, extremism, misinformation, violence, fraud, privacy, or jailbreak behaviour. Hate speech is narrowly defined as offensive language targeting protected groups. | Partial. Dataset reflects Nigerian cultural, political and religious contexts but does not explicitly distinguish global from local harms. | Native speakers | Yes. Annotation performed by native speakers. | Dataset is culturally grounded through native keyword collection and annotation. Paper discusses linguistic characteristics of Hausa, Yoruba and Igbo but does not explicitly address dialect variation. | Not reported. Fleiss' κ was computed and a ≥0.60 acceptance threshold used, but final κ values for Hausa, Yoruba or Igbo are omitted. | Hausa: 4,467 tweets; Yoruba: 4,042 tweets; Igbo: 3,856 tweets. | One of the earliest native-authored offensive language datasets focusing exclusively on three major Nigerian languages. Addresses the scarcity of monolingual Nigerian-language safety datasets. | No dataset release; no license; no reported final IAA; only accuracy reported; class imbalance acknowledged; hate category very small; only Twitter data; keyword-based sampling introduces bias; no per-language precision/recall/F1; no dialect analysis. | Exclude (paper withdrawn from Arxiv) | Not reported | Not publicly available (planned future release) | Twitter/X (Academic API) | Latin | None intentionally. Tweets in other languages were removed during preprocessing. | Offensive – insults, abuses, threatens, mocks or disparages an individual. Hate – offensive language targeting religion, race, gender or political affiliation. Neither – not offensive or hateful. | Three native speakers per language. Keyword validation by three linguistics lecturers. Annotators were trained before annotation. | Partial. Definitions are described in the paper, but no standalone guideline document is released. | XLM-R: Hausa 0.79, Yoruba 0.82, Igbo 0.69. BERT-multilingual: Hausa 0.83, Yoruba 0.83, Igbo 0.87. XLM-T: Hausa 0.81, Yoruba 0.85, Igbo 0.90. AfriBERTa: Hausa 0.85, Yoruba 0.85, Igbo 0.88 (accuracy). |
[HA] LSR (Linguistic Safety Robustness) | 2026 | Hausa (+3) //// Yoruba, Igbo, Igala | Low | https://arxiv.org/pdf/2603.19273 | https://huggingface.co/datasets/Faruna01/lsr-benchmark | Godwin Abuh Faruna | Gemini 2.5 Flash (January 2026 version) | Structured benchmark consisting of paired harmful prompts (English baseline + native-language equivalent). Each record contains language tag, harm category, attack technique, role, pair ID, severity, prompt, and target behaviour. | Native-language prompt engineering. Prompts were manually written to express equivalent harmful intent in culturally appropriate ways rather than translated. English baseline prompts were created separately. | n/a (native) | Physical Harm, Toxicology, Targeted Violence, Historical/Cultural Pretext Framing | No coverage of hate speech, harassment, sexual exploitation, child safety, fraud/scams, cyber abuse, misinformation, self-harm, privacy, political manipulation, biosecurity beyond toxicology, extremism, jailbreak prompts, or social manipulation. Authors acknowledge future expansion to financial fraud, medical misinform... | Strong combination of both. Core harms are globally applicable, while prompts are embedded in West African cultural contexts (Oba, Sarki, Eze, traditional medicine, folklore, historical narratives). | No human annotation in the traditional sense. Responses were automatically classified as Refusal or Compliant using a rule-based refusal classifier. Dataset creation appears to be by the author. | Not formally reported. Prompts were written natively, but the paper explicitly lists lack of native speaker review as a limitation. | Excellent cultural adaptation. Uses culturally authentic authority figures (Oba, Sarki, Eze), traditional medicine, folklore, historical framing, and culturally natural linguistic registers instead of literal translation. Dialect variation itself is not discussed. | Not applicable. No manual annotation task or inter-annotator agreement reported. | 14 probes per language (Hausa, English, Yoruba, Igbo, Igala), totaling approximately 70 benchmark instances. | First benchmark specifically designed to measure cross-lingual safety degradation in West African languages using matched harmful prompts and a novel Refusal Centroid Drift (RCD) metric. Introduces Igala into multilingual AI safety evaluation. | Small benchmark (14 probes/language); evaluates only one model; keyword-based refusal classifier; no human evaluation of model outputs; native speaker validation absent; RCD is an operational proxy rather than a mechanistically validated metric; benchmark author-generated. | Include | Not explicitly stated in paper. Dataset hosted on Hugging Face. | Publicly available (Hugging Face dataset + dashboard + Inspect AI implementation). | Benchmark prompts (synthetic but manually authored). Not collected from social media or web corpora. | Latin | None reported. Each prompt is monolingual. | Physical Harm; Toxicology; Targeted Violence; Historical and Cultural Pretext. Attack techniques: direct, roleplay, historical_pretext, cultural_documentation. Severity labels: HIGH and CRITICAL. | Not applicable. Dataset consists of manually authored prompts; no independent annotation team reported. | No annotation guideline. Evaluation methodology and refusal-classification procedure are fully documented. | Hausa ≈40% (RCD 0.50); English ≈90% refusal (RCD 0.00); Yoruba ≈55% (RCD 0.35); Igbo ≈50% (RCD 0.40); Igala ≈35% (RCD 0.55). |
[HA] AfriSenti (Hausa subset) | 2023 (dataset paper Feb 2023; SemEval task paper Apr 2023). Hausa data originally curated 2022 (NaijaSenti). | 14 African languages incl. hau (also amh, arq, ary, ibo, kin, pcm, orm, pt-MZ, swa, tir, twi, tso, yor). | Low | https://arxiv.org/abs/2302.08956 (dataset paper — primary for judgment columns); arxiv.org/abs/2304.06845 (SemEval-2023 Task 12 paper). | github.com/afrisenti-semeval/afrisent-semeval-2023; HF: masakhane/afrisenti (also HausaNLP/AfriSenti-Twitter). Public. | Muhammad et al. (26 authors) — HausaNLP / Masakhane-linked academic community (Bayero Univ. Kano, Univ. of Pretoria, etc.). | AfroXLMR-large (best baseline and most winning systems), plus per-team variants: AfriBERTa, mBERT, mDeBERTaV3, LaBSE, XLM-T, Bernice, DziriBERT (Arabic dialects only, not relevant to Swahili) — 44 teams total across all languages | prompt-only | Native-authored | n/a (native) | None of the 6 packet harms as labels — sentiment only. Negative-class tweets are safety-adjacent raw material (AfriHate reused AfriSenti negatives for Swahili). | All six missing as explicit labels: hate speech, harassment, self-harm, extremism, sexual content, misinformation. | No | Native speakers | Yes — native review | Monolithic | kappa = 0.66 for Hausa — free-marginal multi-rater kappa (Randolph 2005), reported per language. | 22,155 tweets: train 14,173 / dev 2,678 / test 5,304. Largest slice in the collection. | Native-authored + native-annotated with per-language IAA, exact splits, released individual annotator labels, and public guidelines — a rare fully documented low-resource slice. Quote: 'The tweets were annotated by native speakers.' | (1) Offensive tweets were deleted during collection, so the negative class is sanitized; treating AfriSenti negatives as abusive-content source material is therefore compromised. (2) Latin script only; no Ajami despite the paper discussing it. (3) Residual code-mixed tweets present. (4) License CC BY-NC-SA 4.0; commerc... | Maybe — flag | CC BY 4.0 | Public | Twitter/X (Academic API; location+lexicon heuristics) | Latin | No (Hausa slice; collection notes varieties but doesn't split) | positive / negative / neutral (sentiment — no harm taxonomy) | 3 annotators/tweet; Hausa-speaking NaijaSenti team (Nigeria) | Yes (Mohammad 2016 guidelines; annotator labels released) | Yes — per-language κ=0.66 (Randolph); per-language splits in Table 6 |
[HA] HOC — Hausa Offensive Content dataset ('Detection and Analysis of Offensive Online Content in Hausa Language'). | 2023 (arXiv v1, Nov 2023); revised Mar 2025 (v2); also published at IJCAI AI4G 2024. | hausa only, plus 'Engausa' (Hausa-English code-switched posts) explicitly labeled. | Low | arxiv.org/abs/2311.10541 (also IJCAI AI4G 2024 proceedings). | Not publicly hosted. Request-only — data available on request from corresponding author (Inuwa-Dutse); GitHub repo stated as forthcoming. Flag access status. | Adam, Zandam, Inuwa-Dutse — Federal Univ. Dutse + FUT Babura (Nigeria), Univ. of Huddersfield (UK). Native Hausa-speaking academic team. | Logistic Regression, SVM, Random Forest, XGBoost, MLP, CNN, plus pretrained multilingual transformers XLM-RoBERTa-base and BERT-base-multilingual-cased (mBERT); baseline comparison against Google Translate as a translation-quality benchmark, not a classifier | prompt-only | Native-authored | n/a (native) | Hate speech / harassment territory: offensive and abusive language, distinguishing offensive vs abusive vs banter; religion and politics identified as hotspots. | self-harm, extremism, sexual content, misinformation — none labeled. Hate speech is not a separate label either; everything folds into a single 'offensive' binary. | Yes | Expert | Yes — native review | Monolithic | NONE — no formal inter-annotator agreement conducted; authors cite ethical constraints on exposing annotators to extensive offensive content and defer formal IAA to future work; survey validation used instead. KEY RED FLAG. | Not reported. Verified across arXiv v1 (Nov 2023), v2 (Mar 2025), and the IJCAI AI4G 2024 version: only offensive-term proportions (Table 3) and user-study sample sizes (n=101; n=180) are given. The absence of a reported dataset size is itself a documentation-gap finding; the exact count is obtainable only from the aut... | Native-authored, native-annotated, culturally deep — first offensive-content dataset for Hausa and a strong local-harm exemplar — but size undocumented, no IAA, and gated access. | No IAA; size unreported; request-only access with no stated license; single 'offensive' binary conflates hate/abuse/insult; aggressive preprocessing (posts <8 unique words discarded, lowercasing, stemming) means released data is not raw text; user study skews male (74.7%) and educated, shaping validation judgments. For... | Maybe — flag | Unclear (no stated license) | Request-only (from corresponding author; GitHub forthcoming) | X/Twitter (API keyword search) + Facebook (Hausa pages/groups; Kamus dictionary-guided) | Latin (Boko) | Yes — 'Engausa' Hausa–English code-switched posts explicitly labeled | offensive / non-offensive (+ language, sentiment, topical labels); banter vs abusive distinguished in analysis | Authors (native Hausa speakers) + user-study validation (n=101/180, Nigeria). Precisely: User Study 1, n=180 (Table 1 demographics: 74.7% male / 25.3% female; ages 18–45, mostly 26–35 at 63%; education skews postgraduate 46%); User Study 2, n=101 (Table 4, rating exercise). Both distributed via QuestionPro on Facebook/... | Partial (annotation described; no formal guideline doc) | Yes — mBERT (best overall) Acc/Recall/Prec/F1 = 0.86/0.86/0.89/0.86; XGBoost = 0.86/0.86/0.86/0.86; XLM-RoBERTa-base = 0.84/0.84/0.87/0.84; CNN = 0.84 (unigram-invariant); SVM = 0.82; Random Forest = 0.80; Logistic Regression = 0.80; MLP = 0.80. Trigram feature engineering improved every model except CNN |
[HA] TukaBench | 2026 | Hausa + 6 African (Amharic, Igbo, Chichewa, Swahili, Yorùbá, isiXhosa) + English (8) | Low | https://arxiv.org/abs/2606.01322 | https://huggingface.co/datasets/McGill-NLP/tukabench | Akinode, Li, Hamidouche, Zamir, Becker-Reshef, Adelani (Mila/McGill/Microsoft AI for Good) | - LLM-as-a-Judge: GPT-4.1
- Prompt Validation: GPT-5.2
- Translation & Generation: AfriqueQwen-8B (alongside Google Translate) | prompt-only | Machine-translated | EN→target | physical harm, fraud/deception, malware, expert advice, govt decision-making, sexual/adult, harassment/discrimination, disinformation, privacy, economic harm (JBB 10-cat) || TSPA: Violence (physical harm); Fraud (fraud/deception); Malware; Nudity & Sexual Activity (sexual/adult); Harassment & Bullying + Hateful Content... | None at category level (full JBB taxonomy); real gap is provenance — no natively-authored Hausa prompts | Yes | Native speakers | Yes — native review | Monolithic | Judge–human agreement 69.4% (Hausa); 8 three-way ties/250 | Paper: 986 Hausa prompts (300 JBB-derived + 343 AfriJail-Mono + 343 AfriJail-CS). The current Hugging Face release contains only the three Afri-JBB subsets (300 Hausa prompts); the AfriJail components are not yet uploaded. | Include — native-reviewed, culturally-grounded jailbreak set with a dedicated Hausa slice. | Native review: 2 native Hausa annotators recruited via Upwork, screened for native fluency (TukaBench). | Provenance verification: all African-language prompts were authored in English first, then machine-translated (Google Translate) and post-edited by native speakers — not natively written, contrary to some secondary descriptions. Hausa is the second-best-resourced of the seven African languages and performs comparably t... | Include | CC BY-NC 4.0 (HF card) | LIVE on HF (viewer active) — BUT current release contains only the 3 Afri-JBB subsets (benign/culture/harm, 100/lang each = 300 Hausa prompts). AfriJail-Mono & AfriJail-CS (343+343) NOT yet uploaded. | LLM-authored (English seeds) → African langs; JailbreakBench (JBB) harmful/benign seeds | Latin (Boko/Ajami: Latin) | Yes — AfriJail-CS is Hausa–English code-switched | JBB 10 behaviours: Harassment/Discrimination, Malware/Hacking, Physical harm, Economic harm, Fraud/Deception, Disinformation, Sexual/Adult, Privacy, Expert advice, Government decision-making | 2 native post-editors per language (Hausa); 3 native annotators for response eval | Partial (transcreation + eval protocol described; full guidelines not released) | Yes — Hausa judge–human agreement 69.4%; per-language ASR reported |
[HA] UbuntuGuard | 2026 | Hausa + 9 African (Akan, Ewe, Igbo, Luganda, Nyanja, Swahili, Xhosa, Yoruba, Zulu) + English | Low | https://arxiv.org/abs/2601.12696 | https://github.com/hemhemoh/UbuntuGuard | Abdullahi, Mgonzo, Oduwole, Okewunmi, Owodunni, Singh, Eickhoff (Brown/Ohio State/ML Collective/Tübingen) | - For Dataset & Benchmark Creation: GPT-5, Llama-3.1-405B, Qwen3-235B-a22b
- For Evaluation (15 models in total (8 guardian models and 7 general-purpose open-source models)): NeMoGuard-8B, DeepSeek | multi-turn | Synthetic | EN→target | misinformation/disinformation, hate speech, stereotypes, expert/specialized advice, public-interest (across health, education, legal, politics, culture/religion, finance, labor) | No explicit self-harm / extremism / sexual-content categories — theme+domain taxonomy (5 themes), not the packet's 6 harms | Yes | LLM-generated | Thin / bulk-pass | Monolithic | No IAA (single validator, 20 pairs, calibration only). MT policy score 66.37% — BELOW the 70% threshold | Hausa: 1,656 train + 278 test policy–dialogue pairs (Nigeria; 4 themes, 7 domains) | Include (consider Maybe-flag) — only policy-based multi-turn benchmark covering Hausa; expert-authored, locally-sourced seed queries. | Provenance verification: the 155 domain experts (Google Amplify Initiative) authored the English seed queries only; the paper states query creation is not its contribution. Hausa policies and dialogues are LLM-generated (GPT-5, Llama-3.1-405B, Qwen3-235B) and then Google-translated, with one native validator reviewing ... | Include | Research-only (check repo license) | Public | GPT-5-generated policies + Llama/Qwen dialogues; seed queries from Google Amplify Initiative | Latin (Boko) | No (monolingual Hausa dialogues) | 5 themes: Misinformation & Disinformation, Hate Speech, Stereotypes, Expert/Specialized Advice, Public-Interest harms (× 7 domains) | 1 Hausa native validator on 20 sampled pairs (calibration) | Partial (pipeline documented; per-language guidelines thin) | Yes — per-language PASS/FAIL policy-compliance; Hausa MT quality 66.37% (below 70% bar) |
[SW] UbuntuGuard | 2026 | Swahili + 9 African + English (Kenya context) | Mid | https://arxiv.org/abs/2601.12696 | https://github.com/hemhemoh/UbuntuGuard | Abdullahi, Mgonzo, Oduwole, Okewunmi, Owodunni, Singh, Eickhoff (Brown/Ohio State/ML Collective/Tübingen) | - For Dataset & Benchmark Creation: GPT-5, Llama-3.1-405B, Qwen3-235B-a22b
- For Evaluation (15 models in total (8 guardian models and 7 general-purpose open-source models)): NeMoGuard-8B, DeepSeek | multi-turn | Synthetic | EN→target | misinformation/disinformation, hate speech, stereotypes, expert/specialized advice, public-interest (7 domains) | No explicit self-harm / extremism / sexual-content categories — theme+domain taxonomy, not the 6 packet harms | Yes | LLM-generated | Thin / bulk-pass | Monolithic | No IAA (single validator, calibration only). Swahili MT quality HIGH: transcript 96.99%, policy 93.30% (well above 70%) | Swahili: 1,899 train + 435 test policy–dialogue pairs (Kenya; 5 themes, 6 domains) — largest test slice in the benchmark | Include — Swahili is the best-scoring African slice here; strong MT quality + native calibration. | Same synthetic+MT provenance caveat as the Hausa row: seed queries expert-authored (Amplify), but policies/dialogues LLM-generated & Google-translated. Swahili quality notably higher than Hausa. | Include | Research-only (check repo license) | Public | GPT-5 policies + Llama/Qwen dialogues; Amplify seed queries (Kenya) | Latin | No (monolingual Swahili dialogues) | 5 themes: Misinformation & Disinformation, Hate Speech, Stereotypes, Expert/Specialized Advice, Public-Interest (× 6 domains) | 1 Swahili native validator on 20 pairs (Swahili = 1 of 4 validated langs) | Partial | Yes — Swahili transcript 96.99%, policy 93.30% (well above 70%); largest test slice |
[SW] AfriHate | 2025 | Swahili + 14 other African languages (15 total) | Mid | https://arxiv.org/abs/2501.08284 | https://github.com/AfriHate/AfriHate | Muhammad, Abdulmumin, Ayele, Adelani et al. (Masakhane-linked, 27 authors) | AfroXLMR (continued pretraining + LID adaptation), Aya-23-8B, mBERT, XLM-R, AfriBERTa variants and other multilingual baselines used in benchmark experiments. | prompt-only | Native-authored | n/a (native) | hate speech, harassment/abuse (+ target attributes: ethnicity, politics, gender, disability, religion, other) | self-harm, extremism, sexual content, misinformation not covered — hate/abuse only | Partial | Native speakers | Yes — native review | Monolithic | Free-marginal Randolph κ = 0.55 for Swahili (Table 4; range across collection 0.46–0.81) | 21,092 tweets (Hate 3,974 / Abusive 7,708 / Neutral 9,410); splits 14,760 train / 3,164 dev / 3,168 test — largest slice in AfriHate (Tables 5 & 7) | Include — largest native Swahili hate-speech resource; anchors the native side of the provenance contrast. | Swahili slice re-annotates 3 prior sources (AfriSenti negatives [Muhammad 2023], PolitiKweli misinfo [Amol 2023], Ombui 2019 hate) into AfriHate's 3-class taxonomy — note the provenance chain to avoid double-counting with the corresponding entry/10. | Include | Apache 2.0 (HF card) | Public but GATED on HF (accept conditions); GitHub AfriHate/AfriHate open (data, annotations, lexicons) | Twitter/X | Latin | Some (collection includes Nigerian Pidgin etc.; Swahili slice standard Swahili) | hate, abusive, neutral (+ targets: ethnicity, politics, gender, disability, religion, other) | 5 annotators total, 3 per instance; Swahili native speakers; majority vote (Table 4) | Yes (annotation scheme + lexicons released) | Yes — Swahili is AfriHate's best-performing language: multilingual AfroXLMR-76L macro F1 89.50; GPT-4o 20-shot 84.61 (Tables 8–10) |
[SW] AfriSenti (Swahili subset) | 2023 | Swahili + 13 other African languages (14 total) | Mid | https://arxiv.org/abs/2304.06845 | github.com/afrisenti-semeval/afrisent-semeval-2023 (paper-canonical); live HF: HausaNLP/AfriSenti-Twitter (also shmuhammad/AfriSenti-twitter-sentiment, masakhane/afrisenti mirrors) | Muhammad, Abdulmumin, Ayele et al. (Masakhane / AfriSenti-SemEval) | AfroXLMR-large (best baseline and most winning systems), plus per-team variants: AfriBERTa, mBERT, mDeBERTaV3, LaBSE, XLM-T, Bernice, DziriBERT (Arabic dialects only, not relevant to Swahili) — 44 teams total across all languages | prompt-only | Native-authored | n/a (native) | None directly — sentiment (pos/neg/neutral). Safety-ADJACENT: negative class = source of abusive/sensitive content | All 6 packet harms — this is a sentiment set, not a safety set; included only as a repurposed negative-content source | No | Native speakers | Yes — native review | Monolithic | Not a safety IAA; sentiment agreement reported in AfriSenti paper | 2,401 total — Train 1,198 / Dev 454 / Test 749 (Table 1, this paper) | Maybe-flag — NOT a safety dataset; repurposed negative class only. Flag for the group per inclusion criteria on repurposed data. | Swahili has no manually-curated sentiment lexicon (Table 1 marks Swahili "✗" under Manual Lexicon, unlike Hausa, Igbo, Yorùbá, Amharic, Nigerian Pidgin, Xitsonga, Tigrinya, Oromo which have "✓"). This matters for Task C (zero-shot) since lexicon-based methods were the winning approach there — Swahili wasn't a zero-shot... | Maybe — flag (repurposed data, no IAA, access friction) | CC BY 4.0 | Public | Twitter/X | Latin | No | positive / negative / neutral (SENTIMENT — not a harm taxonomy) | Multiple native annotators (Masakhane); Swahili from Kenya/Tanzania region | Yes (AfriSenti/SemEval-2023 task guidelines public) | Best Task A (monolingual) Swahili F1 = 65.68 (NLNDE, using AfroXLMR-large with language+task adaptive pretraining) vs. AfriSenti's own baseline of 63.40; king001 close second at 64.89. Masakhane-AfriSenti team specifically reported AfroXLMR-base performed better than LaBSE for Swahili (per their system description, §6.... |
[SW] PolitiKweli | 2024 | Swahili + English (code-switched); also pure-English & pure-Swahili subsets | Mid | https://link.springer.com/chapter/10.1007/978-3-031-58495-4_1 | https://github.com/jayneamol/kweli | Amol, Wanzare, Obuhuma (Maseno University, Kenya; SPELLL 2023 / Springer CCIS vol. 2046) | They focused on encoder-based pre-trained language models (PLMs):
- BERT (Bidirectional Encoder Representations from Transformers)
- T5 (Text-to-Text Transfer Transformer)
- SwahBERT | prompt-only | Native-authored | n/a (native) | misinformation/disinformation (political); fake vs fact vs neutral | hate, harassment, self-harm, extremism, sexual content not covered — political misinfo only | Yes | Native speakers | Yes — native review | Multi-variety | Not stated in detail; BERT baseline F1 0.62 | 6,345 Swahili-English code-switched + 22,954 English + 211 Swahili = 29,510 tweets (Swahili-relevant: 6,556) | Include — first Swahili-English code-switched misinfo set; directly covers the misinformation harm + code-switching attack surface. | Bulk of the data is pure English (22,954) — the Swahili-relevant slice is the 6,345 code-switched + 211 Swahili. Don't count the English rows toward Swahili coverage. | Include | Research-only (contact authors; GitHub jayneamol/kweli) | Public (GitHub) | Twitter/X (2022 Kenya general election) | Latin | Yes — Swahili–English code-switched (core of the dataset) | fake / fact / neutral (political misinformation) | Native Kenyan annotators (count not fully specified); Maseno University | Partial (curation & annotation described in chapter) | Yes — BERT baseline F1 0.62 on the Swahili-English CS set |
[SW] Swahili & Code-Switched EN-SW Political Hate Speech | 2025 | Swahili + English-Swahili code-switched | Mid | https://www.sciengine.com/DI/doi/10.3724/2096-7004.di.2025.0053 | Zenodo: doi.org/10.5281/zenodo.14103420; Science Data Bank: scidb.cn/en/c/j00104 (Data Intelligence data repository) | Onyango et al. (Data Intelligence, Chinese Academy of Sciences; CC BY 4.0) | they evaluated the dataset using a domain-specific pre-trained language model:
- SwahBERT
- Bi-LSTM (Bidirectional Long Short-Term Memory)
- SVM (Support Vector Machine) | prompt-only | Native-authored | n/a (native) | hate speech + targets (nationality, social status, politics, disability, ethnicity, gender, religion, other) | self-harm, extremism, sexual content, misinformation not covered — political hate/target only | Partial | Native speakers | Yes — native review | Multi-variety | Randolph free-marginal κ 0.435 (pilot 1) → 0.552 (pilot 2) | 101,014 tweets total (combines PolitiKweli + AfriSenti + Hate_Speech_Kenya, re-annotated with hate/target/language) | Include — adds hate-TARGET + language labels for Swahili/code-switched, which AfriHate/Ombui lacked. | Heavy source overlap: built by MERGING PolitiKweli (the corresponding entry) + AfriSenti (the corresponding entry) + Hate_Speech_Kenya. Same underlying tweets as the corresponding entry re-annotated — major double-count risk; treat as a re-annotation layer, not 101k new instances. | Include | CC BY 4.0 | Public — Zenodo + Science Data Bank (CC BY 4.0) | Twitter/X (combines PolitiKweli + AfriSenti + Hate_Speech_Kenya) | Latin | Yes — Swahili & English-Swahili with explicit language tags; multilingual index 0.72 | hate + targets: nationality, social status, politics, disability, ethnicity, gender, religion, other | 2 pilot rounds; native Swahili annotators (Kenya) | Partial | Reports Randolph κ (0.435→0.552); model baselines limited |
[FR] RTP-LX (RealToxicityPrompts — Language eXpanded) | 2024 (FR released Sept 2023, V0.1) | French + 37 other languages (38 total) | High | https://arxiv.org/abs/2404.14397 | https://github.com/microsoft/RTP-LX | de Wynter et al. (Microsoft; AAAI 2025, pp. 27940–27950) | 1. Frontier Closed Models
GPT-4 Turbo
2. Meta Llama Series
Llama 3 (70B)
Llama 3 (8B)
Llama 2 (70B)
Llama 2 (7B)
Llama Guard (A specialized LLM fine-tuned for content moderation and safety classification)
3. Mistral AI Series
Mistral-v3 (7B)
Mistral-v2 (7B)
4. Google Gemma Series
Gemma (7B)
Gemma (2B) | prompt-only | Transcreated (native speakers) | EN→target | Bias, Identity Attack, Insult, Microaggression, Self-Harm, Sexual Content, Toxicity, Violence (8 ACS-style dims) | misinformation/extremism not explicit dims; harassment folded into toxicity/identity-attack | Yes | Expert | Yes — native review | Multi-variety | Cohen's κw reported; low model–human agreement is a headline finding | ~1,100 French prompts (this slice) + completions; ~30k entries total across all languages | Include — French is in RTP-LX's first release; strongest culturally-annotated French toxicity eval, transcreated not MT. | French is confirmed in RTP-LX's first release (V0.1, Sept 2023: ES/FR/DE/IT/JA/PT/ZH/AR/CS). Language verification against the repository's 38-language list: Swahili was added in V0.3 (Jan 2024) — see the Swahili entry above; Hausa is not covered by RTP-LX. | Include | Research-only (Microsoft; password-protected entries) | Public (GitHub, gated by password) | Reddit-seeded RealToxicityPrompts, professionally transcreated to French | Latin | No (monolingual French; dialects noted where used) | Bias, Identity Attack, Insult, Microaggression, Self-Harm, Sexual Content, Toxicity, Violence | Professional transcreators + multiple human annotators (majority vote) | Partial (harm-category definitions in appendix; full guidelines partial) | Yes — per-language Cohen's κw; French among first-release languages |
[FR] XSafety | 2023 (ACL Findings 2024) | French + 9 others (EN, ZH, ES, BN, AR, HI, RU, JA, DE) — 10 total | High | https://arxiv.org/abs/2310.00905 | https://github.com/Jarviswang94/Multilingual_safety_benchmark | Wang, Tu, Chen, Yuan, Huang, Jiao, Lyu | ChatGPT
PaLM-2 (PaLM2)
LLaMA-2-Chat
Vicuna | prompt-only | Machine-translated | EN→target | 14 safety scenarios: 7 typical (incl. insult, unfairness, crimes, physical/mental health, ethics, privacy) + 6 instruction-attacks + 1 commonsense | Categories are broad-safety, not the packet's exact 6; self-harm present via mental-health, misinformation weakly covered | Partial | Expert | Thin / bulk-pass | Monolithic | Not reported per-language; hybrid create-then-validate | ~2,800 French entries (200 × 14 categories); 28,000 total across 10 languages | Include — French confirmed; broad 14-category safety eval. Apache 2.0. | Provenance is translation from Chinese/English seeds — a good 'translated' contrast against Aya (native) and RTP-LX (transcreated) for the French brief. | Include | Apache 2.0 | Public (GitHub Jarviswang94) | Chinese Safety-Prompts + English SafeText, MT'd then proofread into French | Latin | No | 7 typical (Insult, Unfairness/Discrimination, Crimes/Illegal, Physical Harm, Mental Health, Privacy/Property, Ethics/Morality) + 6 instruction-attacks + 1 commonsense | Professional translators + 2 proofreading rounds (not per-language native-cultural reviewers) | Partial | Yes — unsafe-response rate per language; French reported |
[FR] CultureGuard (Nemotron-Safety-Guard-Dataset-v3) | 2025 | French + 8 (EN, AR, DE, ES, HI, JA, TH, ZH) — v3 later expanded to 12 | High | https://arxiv.org/abs/2508.01710 | https://huggingface.co/datasets/nvidia/Nemotron-Safety-Guard-Dataset-v3 | Joshi, Paul, Singla, Kamath, Evans, Luna, Ghosh et al. (NVIDIA; IJCNLP-AACL 2025) | Llama-3.1-Nemotron-Safety-Guard-8B-v3
Evaluated Safety Guard Models (Baselines): Llama-Nemotron-Safety-Guard-V2, PolyGuard-Qwen, LlamaGuard * ShieldGemma, Granite Guardian, WildGuard, DuoGuard and OmniGuard
Benchmarked Generic LLMs: Llama, Nemotron, Gemma, Mistral, Qwen, ChatGPT, Gemini, and Claude | prompt+response | Synthetic | EN→target | Aegis 2.0 taxonomy: 12 top-level hazard categories + 9 fine-grained (expanded to 23 in v3); incl. hate, sexual, self-harm, violence, weapons, criminal, jailbreak | Broad coverage — few gaps; misinformation lighter than hate/violence | Yes | Synthetic | No | Monolithic | Automated quality filtering (cross-lingual consistency); no per-lang human IAA | ~43k French samples (386,661 total / 9 languages, approx.) | Include — large culturally-adapted French guard-training corpus; strong 'translated+adapted' example. | Distinguish from plain MT: CultureGuard's pipeline adds a cultural-adaptation stage before translation. Source = English Aegis 2.0 / Nemotron-Content-Safety-V2. HF v3 now lists 12 langs (French included). | Include | Open (HF; NVIDIA) | Public (Hugging Face) | Synthetic from English Aegis 2.0 / Nemotron-Content-Safety-V2, culturally adapted → French | Latin | No | Aegis 2.0: 12 top-level hazard categories + 9 fine-grained (v3: 23 categories) incl. jailbreak | Synthetic (LLM-generated); no per-language human annotators; cross-lingual consistency filter | Yes (pipeline & taxonomy documented; Aegis 2.0 public) | Yes — per-language guard-model F1; French reported vs baselines |
[FR] Jigsaw Multilingual Toxic Comment Classification | 2020 | French + Italian, Portuguese, Russian, Spanish, Turkish (6 test langs; train EN-only) | High | https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification | https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/data | Jigsaw / Google Conversation AI | Model used is unknown | prompt-only | Native-authored | n/a (native) | toxicity (binary) + subtypes: severe toxic, obscene, threat, insult, identity hate | self-harm, extremism, misinformation not covered — general toxicity/hate only | No | Crowdsourced | No | Monolithic | Multiple annotators per comment (up to 10); PerspectiveAPI-aligned | French: 10,920 test comments (+ shared validation pool); no French training data | Include — rare NON-social-media French toxicity source (Wikipedia talk pages); organic native French text. | LINK-CHECK: the arXiv ID in the source table (2010.12474) is NOT the Jigsaw dataset — Jigsaw is a Kaggle competition with no matching arXiv paper. Point the paper column at the Kaggle page, not that arXiv link. French is eval-only (no train split). | Include | CC BY-NC-SA 4.0 (underlying); Kaggle competition terms | Public (Kaggle, free account) | Wikipedia talk-page comments (organic French) | Latin | No | toxic, severe toxic, obscene, threat, insult, identity hate (PerspectiveAPI-aligned) | Crowdsourced (up to 10 raters/comment); Civil Comments / Wikipedia — region not language-targeted | Partial (PerspectiveAPI definitions public; per-item guidelines partial) | Yes — French test AUC/F1 widely benchmarked (10,920 French test comments) |
[SW] MultiJail | 2023 | Swahili + 8 others (ZH, IT, VI, AR, KO, TH, BN, JV) + English (10 total) | Mid | https://arxiv.org/abs/2310.06474 | https://github.com/DAMO-NLP-SG/multilingual-safety-for-LLMs | Deng, Zhang, Pan, Bing (DAMO Academy / NTU) | ChatGPT (GPT-3.5-turbo-0613), GPT-4 (GPT-4-0613) as primary; Llama2-7b-chat, Vicuna-7b-v1.5, SeaLLM-7B-v2 as supplementary open-source comparisons | prompt-only | Human-translated by native speakers (main dataset); machine-translation via Google Translate used only as a separate ablation, not the primary dataset | EN→Swahili | 18-category tag taxonomy (Figure 7/Appendix A.3, inherited from Anthropic's red-teaming tags): Violence & incitement (18.8%), Discrimination & injustice (11.3%), Hate speech & offensive language (8.4%), Bullying & harassment (7.5%), Non-violent unethical behavior (6.6%), Conspiracy theories & misinformation (6.4%), The... | No standalone "extremism" label (folded into Violence & incitement / Terrorism & organized crime) | No — the taxonomy and prompts are generically applied across languages with no discussion of culturally-specific vs universal harm | Native-speaker translators for construction; human evaluators (English-fluent) + GPT-4-as-judge for safety labeling of model outputs | Yes — native Swahili speakers translated; separate native-speaker group verified translation quality | Monolithic — no dialect variation noted for Swahili specifically | Two distinct scores, neither Swahili-specific: (1) translation-quality verification targeted a >97% pass rate across all 9 languages (not a kappa); (2) Cohen's κ = 0.86 between human evaluators and GPT-4-as-safety-judge (measuring evaluator agreement, not translation/annotation IAA) — reported at the aggregate level on... | 315 Swahili prompts (315 English seeds translated into each of 9 langs; 3,150 total across 10 languages) | Native-translated, quality-verified, richly tagged (18 categories) jailbreak set with strong, specific Swahili results across both scenarios and multiple models — one of the more rigorous native-review entries in the sheet | (1) The 300 red-teaming-sourced prompts come from Anthropic's red-teaming dataset (Ganguli et al. 2022) specifically — distinct from HH-RLHF — worth citing precisely to avoid confusion; (2) safety evaluation itself relies on Google-Translate-then-GPT-4-judge, introducing a translation-noise layer on the output side tha... | Include | Not explicitly stated in the repo per available documentation — flagged for direct verification (other DAMO-NLP-SG repos commonly use MIT or BSD-3-Clause, but this specific repo's license wasn't confirmed in search results) | Public (GitHub) | 15 curated prompts from the GPT-4 system card (OpenAI, 2023) + 300 sampled from Anthropic's red-teaming dataset (Ganguli et al., 2022), human-translated to Swahili | Latin | No — monolingual Swahili prompts | Safe / Unsafe / Invalid (response classification) + 18 safety tags: Violence & incitement, Discrimination & injustice, Hate speech & offensive language, Bullying & harassment, Non-violent unethical behavior, Conspiracy theories & misinformation, Theft, Weapons, Adult content, Substance abuse & banned substances, Terror... | Native-speaker translators + separate native verifiers (>97% pass-rate target) | Partial — GPT-4 evaluation prompt is published verbatim (Appendix A.2); translator instructions are described narratively but no standalone translator guideline doc is released | Extensive, Swahili-specific, across both models and scenarios: Unintentional — ChatGPT: 7.94% unsafe / 91.75% safe / 0.32% invalid; GPT-4: 6.35% unsafe / 92.06% safe / 1.59% invalid. Intentional — ChatGPT: 83.49% unsafe / 15.56% safe / 0.95% invalid; GPT-4: 43.49% unsafe / 56.51% safe / 0.00% invalid. Open-source compa... |
[SW] XTREMESPEECH | 2022 | Swahili + 3 (Brazilian Portuguese, German, Hindi) | Mid | https://arxiv.org/pdf/2203.11764 | https://github.com/antmarakis/xtremespeech | Antonis Maronikolakis, Axel Wisiorek, Leah Nann, Haris Jabbar, Sahana Udupa, Hinrich Schütze (LMU Munich) | mBERT, XLM-R, monolingual langBERT (bert-base-uncased-swahili for Kenya), SVM, LSTM baselines | prompt-only | Native-authored / organically collected — fact-checkers sourced real posts from Facebook, Twitter, WhatsApp as encountered, not queried or translated | n/a (native) | Custom 3-way "extreme speech" scale: derogatory (moderation-level), exclusionary (incites discrimination), dangerous (incites violence) + target-group labels. Maps loosely onto our hate speech / harassment categories (dangerous ≈ incitement/extremism-adjacent) | Self-harm, sexual content, misinformation — not covered. No explicit "extremism" label either; folded into "dangerous" | Yes — core design goal. Kenya-specific example: the Kikuyu/Kalenjin power-dynamics disagreement led to a new locally-necessary label ("large ethnic groups") rather than forcing data into a Western minority/majority frame | Independent, accredited local fact-checkers (not company/gov employees) — both collected and labeled data | Yes — Kenyan annotators, native to the region; anthropology team vetted their local-context expertise | Multi-variety | Kenya-specific (Table 14): κ = 0.13, α = 0.21, ICC(3,k) = 0.47, Target κ = 0.50; accuracy 58.1% overall (Derogatory 69.4%, Exclusionary 11.8%, Dangerous 57.1%); M/R accuracy 69.4%/43.0% | Kenya-specific (Table 14): κ = 0.13, α = 0.21, ICC(3,k) = 0.47, Target κ = 0.50; accuracy 58.1% overall (Derogatory 69.4%, Exclusionary 11.8%, Dangerous 57.1%); M/R accuracy 69.4%/43.0% | Genuinely native, community-embedded collection with a documented local-harm finding (ethnic-group relabeling); but pure-Swahili text is a small minority of the Kenya slice (405 of 5,181), and IAA for the exclusionary/dangerous distinction is weak | (1) IAA is low by any standard — Kenya κ=0.13 is "poor," even below the paper's own already-low overall average (0.23); (2) most of the "Kenya" data is English or code-switched, not monolingual Swahili — pure-Swahili is only ~8% of the slice; (3) no neutral/negative class was collected at all — 100% of passages are som... | Include — flag low IAA and small monolingual-Swahili share | Not stated in the paper; GitHub repo contains code only, no LICENSE file governing the data itself | Request-only | Facebook, Twitter, WhatsApp (Kenya); platform-agnostic, collected Q3 2020–end 2021 | Latin | Yes — explicitly tagged category ("Both"); 2,081 of 5,181 Kenya passages (~40%) are code-switched, plus another ~52% pure English | Derogatory Extreme Speech / Exclusionary Extreme Speech / Dangerous Extreme Speech; Targets: ethnic minorities, immigrants, religious minorities, sexual minorities, women, racialized groups, historically oppressed caste groups, indigenous groups, large ethnic groups (added), politicians, legacy media, the state | At least 2 accredited fact-checkers for Kenya (0 female / 2 male per the data statement); interviewed and vetted by the paper's anthropology team for local-context expertise | Partial — definitions given in-paper and shared with annotators as instructions, but no standalone guideline document released | Kenya-specific F1 reported across all tasks: EXTREMITY (Table 5/16), REMOVAL (Table 6/17), TARGET LRAP (Table 7/18), plus cross-country transfer (Tables 8–9, 19–20) and monolingual English-only Kenya (KEen) transfer results |
[SW] Kiswahili Deep Dive, from "International Joint Testing Exercise: Agentic Testing — Advancing Methodologies for Agentic Evaluations Across Domains" (3rd Joint Testing Exercise) | 2025 | Swahili (Kiswahili) + 8 (English (source) → Farsi, French, Hindi, Japanese, Korean, Mandarin Chinese, Telugu (8 target languages); Kiswahili is the Kenya-contributed slice) | Mid | https://arxiv.org/pdf/2601.15679 | None — this PDF is an evaluation report, not a released dataset repo. Component benchmarks it draws on are separately public (e.g. huggingface.co/datasets/ScaleAI/BrowserART, AgentHarm/InjecAgent/ToolEmu on arXiv/GitHub), but the translated Kiswahili task/tool files themselves are not released | No individual byline; credited to "AI Safety Institutes & Government-Mandated Offices from Singapore, Japan, Australia, Canada, European Commission, France, Kenya, South Korea, and UK AI Security Institute" — Kiswahili section specifically "Contributed by Kenya AISI" | Agents: Model A (larger, closed-weight) & Model B (smaller, open-weight); Judges: Model C (large, closed-weight) & Model D (smaller, open-weight, English-only) — all anonymized in the report | Multi-turn agentic trajectories (task prompt + tool calls with reasoning + final output) | Mixed: most tasks machine-translated from English then human-validated by native speakers (per general Methodology section); the Kenya-contributed fraud task set (10 tasks) was natively authored in Kiswahili context, not translated | EN→Swahili for the bulk of tasks (AgentDojo/AgentHarm/Agent-SafetyBench/BrowserART/HarmBench/InjecAgent/ToolEmu-derived); n/a (native) for the Kenya-authored fraud subset | Fraud (financial fraud, identity theft, unauthorized access, phishing, impersonation) and Sensitive Information Leakage (chat history, bank details, passwords, PII) | hate speech, harassment, self-harm, extremism, sexual content, misinformation are not addressed anywhere in this report. This dataset is entirely outside that taxonomy (agentic fraud/privacy risk instead) | Partial — the Kenya-contributed fraud tasks were locally authored with Kenya-specific context (e.g. referencing Safaricom/M-Pesa, later swapped for Airtel/UPI equivalents when reused for the Hindi slice), but the bulk of the Kiswahili task set is a straight translation of generic English benchmark tasks with no cultura... | Kenya AISI staff (native-speaking human annotators) as primary ground truth; Model C and Model D as secondary judge-LLMs for comparison | Yes — Kenya AISI is explicitly credited with providing "native fluency and linguistic expertise" and authored/validated the Kiswahili section | Minimal — no dialect discussion; cultural adaptation of translated tools/tasks is explicitly flagged in the general Limitations section as "not comprehensive" | No human-human IAA reported — this exercise uses human-vs-judge-LLM discrepancy rate instead of inter-annotator kappa. Kiswahili discrepancy: Model D 41.6% overall (Fraud 37.9%, Sensitive info 45.2%), Model C 37.7% overall (Fraud 30.1%, Sensitive info 45.2%) — methodologically different from the kappa/alpha scores used... | 156 tasks total (77 fraud + 79 sensitive-info-leakage), each run against both agent models; ~132 corresponding tools | Genuinely native-reviewed with a locally-authored task component, but the harm domain (fraud/agentic misuse) doesn't overlap with your project's tracked categories at all, the sample is very small (authors' own words: "difficult to draw statistically significant... findings"), and no dataset is actually released | (1) Authors explicitly flag the 156-task sample as too small for statistically significant conclusions; (2) no human-human IAA — relies on human-vs-LLM-judge discrepancy as the reliability proxy, a different construct than kappa; (3) heavy reliance on machine translation "as a first cut," with code/tool translation spe... | Maybe — flag. Flag reason: no public dataset artifact, despite genuine native review and Kenya-authored content | Not stated for the report or the translated task set. Underlying component benchmarks (AgentHarm, InjecAgent, ToolEmu, BrowserART) carry their own separate research licenses | Public — freely downloadable PDF report (arXiv:2601.15679; also hosted at sgaisi.sg). Not public: the actual Kiswahili-translated tasks/tools themselves are not released as a separate downloadable dataset | Synthetic agentic task/tool environments adapted from AgentDojo, AgentHarm, Agent-SafetyBench, BrowserART, HarmBench, InjecAgent, ToolEmu + an original Kenya-authored fraud task set (phishing, impersonation) — not organic social media text | Latin | Not a designed feature of the dataset — no code-switched task variants were built. However, English code-switching appears as an emergent model behavior (Model B defaulting to English mid-response) rather than an annotated input property | Pass/Fail scenarios per risk scenario: Malicious user task; Benign user task + maliciously injected instruction; Benign user task (underspecified/unsafe potential) — crossed with two risk categories (Fraud; Sensitive Information Leakage); secondary qualitative metrics: Linguistic Fidelity, Comprehensibility, Absence of... | Kenya AISI staff (native Kiswahili speakers, government AI safety institute employees); each participating AISI annotated 1–3 languages; exact Kenya headcount not disclosed | Partial — pass/fail scenario definitions and metric definitions are published in the report itself (Tables 6–7), but the full internal annotation guide/rubric used by AISIs is not released as a standalone document | Yes, extensive Kiswahili-specific tables: pass rate by model & risk type (Table 16), pass rate by risk scenario (Table 17), qualitative metrics (comprehensibility 100%/100%, linguistic fidelity 100%/74%, hallucination 2%/1%, logical consistency 97%/98%), and judge-LLM discrepancy rates (Table 18) |
[SW] Marx & Dunaiski — Multilingual Jailbreaking Using Low-Resource Languages (Kiswahili slice) | 2026 | Swahili (Kiswahili) + (Afrikaans, isiXhosa, isiZulu (+ English baseline); source material drawn from MultiJail (originally 10 languages) and MHJ (English only)) | Mid | arXiv:2605.18239v1 | None found — no code/data availability statement in the paper | Dylan Marx, Marcel Dunaiski | Preliminary: Gemini-2.0-flash-lite, GPT-4o-mini. Main automated: + Gemini-2.0-flash, GPT-4o, Claude-3.5-Haiku, DeepSeek-V3, Grok-3-mini (7 total). Human red-teaming: only GPT-4o, Gemini-2.0-flash, Grok (3 of the 7) | Mixed — single-turn (from MultiJail) and multi-turn (from MHJ); multi-turn is the more effective and more thoroughly tested condition | Machine-translated (Google Translate) as the primary method; a filtered subset then human-refined/extended via live red-teaming conversation (not full re-translation — humans corrected only the first prompt, then improvised subsequent turns toward the same goal) | EN→Kiswahili | Custom 5-category intent taxonomy (Table 5/Appendix A): Misinformation, Harassment or Hate speech, Dangerous (encouraging physical harm), Illegal Activities, Exploitation and Abuse — maps onto hate/harassment and misinformation from our 6-harm framework directly | No standalone self-harm category (folded into "Dangerous"); no standalone sexual-content category (folded into "Exploitation and Abuse"); no explicit extremism category | No — framed entirely around resource-level/translation-quality effects, not cultural or local-harm distinctions | Translation quality: automated metrics only (BERTScore/METEOR/BLEU), no humans. Response/harm labeling: the authors themselves ("we manually reviewed and annotated all LLMs' responses" — count/background unspecified). Human red-teaming: native speakers, fluent in English | Yes, but thin — only one native Kiswahili speaker conducted red-teaming, versus three each for English, Afrikaans, isiXhosa, and isiZulu; the paper itself flags this as limiting the reliability of the Kiswahili human-vs-automated comparison | None discussed — monolithic Kiswahili, no dialect variation | No human-human IAA/kappa reported anywhere. Translation-quality proxies for Kiswahili (Table 1): BERTScore 0.91, METEOR 0.79, BLEU 0.52 (mean); quality thresholds applied (Table 2): BERTScore 0.94, METEOR 0.88, BLEU 0.64 | Not clearly broken out by language — the paper reports only combined totals: 47 high-quality-filtered MultiJail prompts + 55 high-quality-filtered MHJ conversations across all 4 languages together. The one Kiswahili-specific count available is from the human red-teaming table: N=10 conversations (across Gemini/GPT-4o/G... | Valuable as a temporal check on whether 2023-era Swahili jailbreak vulnerability (MultiJail's 83.49%) persists in 2026 models — it mostly doesn't for single-turn, but multi-turn remains highly effective; however, the Kiswahili-specific evidence base is thin (1 translator, ~10 red-teamed conversations, no clear per-lang... | (1) Kiswahili human red-teaming rests on a single translator — paper explicitly notes "this comparison is based on a single translator," unlike the 3-per-language design elsewhere; (2) per-language breakdown of the 47+55 filtered prompts is not given in the main text — a real documentation gap; (3) heavy quality filter... | Maybe — flag. Flag reasons: single-translator Kiswahili sample, unclear per-language size, and direct content overlap with an already-included MultiJail row | Not stated | Not public — no dataset repository or download link found in the paper | Derived from MultiJail (Deng et al. 2024) and MHJ (Li et al. 2024) benchmark prompts, Google-Translate-translated, partially human-refined through live red-teaming conversations with GPT-4o, Gemini-2.0-flash, and Grok | Latin | No — translations are monolingual Kiswahili text; quality assessed via back-translation to English, not code-switching | Response types: Safe / Harmful / Unrelated-Invalid; intent categories (Table 5): Misinformation, Harassment or Hate speech, Dangerous, Illegal Activities, Exploitation and Abuse | Red-teaming: 1 native Kiswahili speaker (fluent in English) — versus 3 for other languages. Response labeling: unspecified count/background, described only as "we manually reviewed and annotated" | No — category definitions appear only in Table 5 of the paper itself; no standalone guideline document | Extensive Kiswahili-specific results: Single-turn unsafe rate (Fig. 1) 4.3–13.0% across 4 models; Multi-turn automated unsafe rate (Fig. 2) 41.8–70.9% across 7 models; Human red-teaming (Table 4) 62.34%→63.33% (automated vs. human, +0.99%); model-specific human red-teaming (Table 6): Gemini 50.0% harmful, GPT-4o 70.0%,... |
[FR] PolyGuardMix (training) / PolyGuardPrompts (eval) - fr slice | 2025 (arXiv Apr 2025, v2 Aug 2025). | 17 languages: ar, zh, cs, nl, en, fr, de, hi, th, it, ja, ko, pl, pt, ru, es, sv. | High (French is a high-resource language within this set; the paper notes the methodology does NOT extend to low-resource languages). | arxiv.org/abs/2504.04377 | huggingface.co/datasets/ToxicityPrompts/PolyGuardPrompts; PGMix training corpus on the same HF collection; code at github.com/kpriyanshu256/polyguard. | Kumar, Jain, Yerukola, Jiang, Beniwal, Hartvigsen, Sap - Carnegie Mellon, Univ. of Washington, IIT Gandhinagar, Univ. of Virginia, Allen Institute for AI. | PolyGuard Qwen2.5 (fine-tuned Qwen2.5-7B-Instruct), PolyGuard Ministral (fine-tuned Ministral-8B-Instruct-2410), PolyGuard Smol (fine-tuned Qwen2.5-0.5B-Instruct) - trained via LoRA on PGMix. Baselines evaluated: Llama-Guard-2, Llama-Guard-3-8B, Aegis-Defensive, MD-Judge, DuoGuard, Perspective API, OpenAI Omni Moderati... | Prompt-output pairs with labels for prompt harmfulness, response harmfulness, and response refusal, plus category-violation lists (multi-label). PGMix=1.91M training pairs; PGPrompts=29K eval pairs. | MIXED for French: PGMix combines (a) machine-translated WildGuardMix (English source, translated via NLLB-3.3B / TowerInstruct-7B) and (b) naturally-occurring In-The-Wild samples (LMSys-Chat-1M, WildChat) that are natively multilingual, not translated. PGPrompts test split translated via a GPT-4o agentic pipeline. | English -> French (machine translation) for the WildGuardMix-derived portion; native/no-translation for the ITW portion. The two are blended, not separated in the release. | Follows the MLCommons Safety Taxonomy (aligned to Llama-Guard-3's 14 categories: Violent Crimes, Non-Violent Crimes, Sex Crimes, Child Exploitation, Defamation, Specialized Advice, Privacy, IP, Indiscriminate Weapons, Hate, Self-Harm, Sexual Content, Elections, Code Interpreter Abuse). TSPA mapping: Hate->Hateful Conte... | No dedicated Extremism/Dangerous-Organizations category distinct from generic 'crimes'; no misinformation category beyond 'Elections'. | Global - taxonomy is a single US-centric standard (MLCommons/Llama-Guard-3) propagated to all 17 languages via translated English labels; not locally adapted for French-specific harm norms. | AUTOMATED, not human, for the primary labels: Llama-Guard-3-8B + GPT-4o as dual LLM-judges on the English source, with Llama-3.1-405B-Instruct breaking ties; labels then PROPAGATED (not re-annotated) to French translations. ITW samples labeled by GPT-4o only (Llama-Guard-3 performs poorly multilingually). | No - NOT a native review for French. Human validation exists but is limited to 50 sampled datapoints for French (out of the full French slice), 3 annotators each, recruited via Prolific and screened for French fluency (not necessarily native). This is a spot-check, not native annotation of the corpus. | Monolithic; no dialect handling (e.g. no distinction for Quebecois vs metropolitan French); translation-based data risks losing register/slang. | French translation quality: DA+SQM score 82.12/100 (highest of all 16 non-English languages). Source-safety-label Krippendorff's alpha=0.48; target(French)-safety-label alpha=0.47; source-target label agreement alpha=1.00 (perfect, highest of all languages). | Not stated precisely for French alone in the paper; French is one of 17 roughly-equal-sized slices of 1.91M (PGMix) and 29K (PGPrompts) - implies roughly ~112K-115K PGMix and ~1.7K PGPrompts for French, but exact per-language counts are not tabulated in-text. | High-quality translation (best DA+SQM score among all languages) and perfect source-target label agreement, but labels are LLM-generated and only translation-propagated, not independently annotated for French; human validation is a 50-sample spot check, not full native review. | Machine-translation-introduced errors possible despite high scores; labels inherit English-centric MLCommons taxonomy; ITW portion (WildChat/LMSys) may contain PII since it's real user chat logs; dataset intentionally NOT extended to low-resource languages (equity gap by design). | Maybe - flag. Largest safety training corpus and strong translation-quality metrics for French specifically, but labels are automated/propagated rather than natively annotated, and human validation is a thin spot-check. | Released under the ODC-BY license (Open Data Commons Attribution). | Public - live on Hugging Face and GitHub (verified). | WildGuardMix (translated) + LMSys-Chat-1M and WildChat (naturally-occurring human-LLM chat logs). | Latin (standard French orthography). | Not explicitly tracked/labeled as a dimension; ITW chat logs may contain organic code-switching but it is not annotated as such. | MLCommons Safety Taxonomy categories (S1-S14, mirrors Llama-Guard-3): Violent Crimes, Non-Violent Crimes, Sex Crimes, Child Exploitation, Defamation, Specialized Advice, Privacy, Intellectual Property, Indiscriminate Weapons, Hate, Self-Harm, Sexual Content, Elections, Code Interpreter Abuse. | Prolific-recruited annotators, pre-screened for English + target-language fluency (not confirmed native), 90-100% approval rate; 3 annotators per datapoint; 191 unique annotators across 16 languages and 24 countries (110 male/81 female; 84 White/79 Black/12 Mixed/10 Asian/5 Other) - demographics are aggregate across AL... | Partial - annotation framework/instructions shown in paper figures (consent, instructions, DA+SQM scale) but no standalone published guideline document specific to French. | Not language-specific in the main results tables (Tables 1-3 report aggregate/English scores); the only French-specific number reported is the Aya Red-teaming recall comparison: PG Qwen2.5 recall 0.916 in French vs. 0.706 for a translate-then-classify baseline (Llama-Guard-3-8B via TowerInstruct translation). |
[FR] PolygloToxicityPrompts (PTP) - fr slice | 2024 (arXiv May 2024, v3 Aug 2024). | 17 languages: ar, zh, cs, nl, en, fr, de, hi, id, it, ja, ko, pl, pt, ru, es, sv. | High (French is one of the higher-resource languages in the set, though the paper notes French had a very low natural toxicity yield before augmentation). | arxiv.org/abs/2405.09373 | huggingface.co/datasets (ToxicityPrompts org); code at github.com/kpriyanshu256/polyglo-toxicity-prompts. | Jain, Kumar, Gehman, Zhou, Hartvigsen, Sap - Carnegie Mellon University, University of Virginia, Allen Institute for AI. | 62 LLMs benchmarked on the full PTP_SMALL run (Llama-2 7B/13B/70B base+chat, Pythia suite, Mistral-7B variants, GPT-3.5-Turbo (0301), Aya101, Bloomz family, Qwen-7B-Chat, Yi-6B-Chat, Gemma variants, Tulu-2 family, OLMo-7B-Instruct, Swallow suite, etc.) - evaluated as base/instruct/preference-tuned toxicity generators, ... | 25K naturally-occurring document-level prompts per language (document split in half at the character level to form prompt + continuation), each scored for toxicity by Perspective API across 6 attributes (TOXICITY, INSULT, THREAT, PROFANITY, IDENTITY_ATTACK, SEVERE_TOXICITY). Not harm-category-labeled; purely a toxicity... | MOSTLY native - scraped from mC4 and The Pile (real French web text), stratified into 4 toxicity bands via Perspective API scores. A minority (~16.8% dataset-wide, concentrated in low-toxicity-yield languages including French) is SYNTHETIC: toxic mC4/Pile samples machine-translated into French via NLLB-3.3B to compensa... | n/a for the majority native portion; English -> French (NLLB-3.3B) for the synthetic augmentation subset specific to French (French is explicitly named among the 9 low-toxicity-yield languages needing this augmentation). | None of the 6 TSPA-tracked types directly - this is a continuous TOXICITY score (Perspective API categories: toxicity, insult, threat, profanity, identity attack, severe toxicity), not a harm-type classification. Identity Attack loosely maps to Hateful Content; Threat loosely maps to Violence. | No Harassment/Bullying, Self-Harm, Sexual-Content, Misinformation, or Extremism/Organizations labels - Perspective API's 6 attributes are a narrower, orthogonal axis. | Global - Perspective API is a single fixed multilingual classifier (not locally recalibrated per language/culture); the paper itself cites known Perspective API language bias (flags German specifically; implies unverified bias risk for other languages including French). | No human annotation of harm at all - 100% automated scoring via Perspective API (industry tool), described by the authors' own Limitations section as a validity gap ('human validations... would help corroborate our results... but... make validations challenging'). | No - explicitly NOT natively reviewed. The paper's own Limitations state no human validation of toxicity judgments was performed for any language, French included, due to scale and cost. | Monolithic; no French-dialect handling; mC4 filters remove pages with 'bad words' lists which may itself embed English-centric or incomplete profanity coverage for French. | None - no inter-annotator agreement exists because there is no human annotation; Perspective API is treated as ground truth. | 25,000 naturally-occurring French prompts (part of the 425K total across 17 languages) + PTP_SMALL stratified subsample of 5,000 French prompts used for the actual 62-model benchmarking runs. | Naturally-occurring, large-scale (25K), stratified by toxicity level, and part of a well-cited multilingual benchmark - but zero human validation of the toxicity labels and partial reliance on MT-augmented synthetic toxicity for French specifically. | Toxicity is Perspective-API-only with no human check (self-acknowledged limitation); mC4's bad-words filtering suppressed the natural French toxicity rate below 0.01%, forcing synthetic augmentation; prompts are much longer than earlier benchmarks (~400 GPT-4 tokens) which may not reflect realistic user-prompt length. | Maybe - flag. Valuable as a naturally-occurring, large-scale toxicity stress-test, but not a harm-category-labeled safety dataset, has zero human validation, and its French toxic examples are partly MT-synthetic. | AI2 ImpACT License - Low Risk Artifacts (a dual-use mitigation license, not a standard open license). | Public - live on Hugging Face and GitHub (verified). | mC4 and The Pile web-text corpora (blogs, news, travel, hosting sites per metadata analysis). | Latin (standard French orthography). | Not tracked/labeled; web-scraped documents may contain incidental code-switching but this is not annotated. | No harm taxonomy - continuous Perspective API scores: TOXICITY, SEVERE_TOXICITY, INSULT, THREAT, PROFANITY, IDENTITY_ATTACK (0-1 scale each). | n/a - no human annotators; automated Perspective-API-only pipeline. | n/a - no annotation guidelines exist since there is no human annotation. | Not broken out specifically for French in the main tables; results are reported by LANGUAGE RESOURCE TIER (high/medium/low) in Table 2, and French falls in the 'high' resource bucket grouped with de/en/es. Table 5 (Appendix D) groups French under 'high - de, en, es, fr' for several model families' AT/EMT/Empirical Prob... |
[FR] Ubisoft Multilingual Game-Chat Toxicity system (ToxBuster / unified soft-prompting model) - fr slice (reference, not a released dataset) | 2025 (arXiv 2506.06347, accepted KDD 2025 v.2); builds on earlier ToxBuster work (2023-2024). | Evaluation covers French, German, Portuguese, Russian (4 languages tested on real game-chat); label-transfer framework extends to 7 additional languages. | High (French is a well-resourced language; low-resource status is not the issue here - proprietary/production data access is). | arxiv.org/abs/2506.06347 | Not public. Underlying human-annotated French hate-speech source dataset (Ousidhoum et al. 2019) used for label transfer is third-party academic data; the game-chat data itself (Rainbow Six Siege / For Honor) is proprietary production data. | Ubisoft La Forge / Ubisoft Data Office research team; academic collaboration referenced with a university 'Complex Data Lab'. | GPT-4o-mini used as the LLM label-transfer engine (not the safety classifier itself); the deployed detection model is BERT-base (ToxBuster architecture). | Real-time production toxicity CLASSIFIER (BERT-based, ToxBuster architecture with soft-prompting game-context tokens), not a static released dataset - output is a sanctionable/non-sanctionable toxicity flag per chat message. | HYBRID / label-transfer: the French evaluation set is built by taking an existing human-annotated French hate-speech dataset (Ousidhoum et al. 2019) and using GPT-4o-mini to transfer/generate labels onto Ubisoft's own real in-game French chat text, validated by a consistency filter - not directly human-annotated game c... | n/a for direct translation; but methodologically a LABEL-TRANSFER from a general French hate-speech corpus onto game-chat domain text via GPT-4o-mini, a domain-transfer analogous to translation risk. | Hate speech / toxicity broadly - maps to TSPA Hateful Content and Harassment and Bullying (game-chat harassment, insults, slurs); the underlying French source corpus (Ousidhoum et al.) is a hate-speech dataset specifically. | No coverage of Self-Harm, Sexual Content, Extremism, Misinformation, or general Violence - scope is narrowly toxicity/hate/harassment in gaming-chat context. | Local-ish for the underlying source corpus (a dedicated French hate-speech dataset), but the GAME-CHAT-SPECIFIC French labels are machine-generated via GPT-4o-mini label transfer, not natively re-annotated for gaming register/slang. | No human annotation of the Ubisoft French game-chat data itself - labels are GPT-4o-mini-generated via the label-transfer framework, validated by an automated consistency filter. | No, for the game-chat-specific data - human annotation exists only in the SOURCE hate-speech dataset (Ousidhoum et al. 2019), not in the Ubisoft game-chat text itself. | Domain-specific: the system is explicitly designed to distinguish 'game-specific language' (jargon, slang) from genuinely harmful language - a real strength for gaming register, but not general dialect/cultural adaptation. | No IAA reported in available excerpts for the French evaluation; performance is reported as macro F1 (32.96%-58.88% range across the four languages) rather than an agreement statistic. | Not stated in a headline number; 'real game chat data' evaluation across French/German/Portuguese/Russian with macro F1 32.96%-58.88%; exact French sample count not found - flag as not reported. | Real production-scale relevance and genuine gaming-register content - the strongest ecological validity in this batch - but the dataset itself is proprietary/inaccessible, and French labels are machine-generated via label transfer. | Not a public dataset; label-transfer methodology (GPT-4o-mini) inherits whatever biases/errors that model has; macro F1 for French not explicitly isolated (reported as a range across 4 languages). | Exclude (as a re-usable dataset) - flag as a methodology/baseline reference only. Cannot be included in a training-data catalogue since underlying French text/labels are not obtainable. | Not stated / not applicable - proprietary Ubisoft production data; no public license. | Not public - proprietary production data; the underlying academic French hate-speech source dataset has its own separate license (not covered in these excerpts). | In-game chat logs from Ubisoft titles (Rainbow Six Siege, For Honor) + external academic hate-speech datasets used for label transfer. | Latin (game-chat French, likely including gaming shorthand per the 'slang, abbreviations, coded language' challenge noted in Ubisoft's own description). | Not specifically discussed for French; the broader ToxBuster program analyzed language distribution across 12 languages including French using Lingua-py. | No published harm taxonomy for the French slice; French is treated as a binary/scalar toxicity-sanctionable classification, not category-labeled. | Not disclosed for the French label-transfer step; underlying Ousidhoum et al. 2019 source dataset's own annotator details not included in these excerpts. | Not stated / not found in available excerpts. | Yes, at a coarse level: macro F1 32.96%-58.88% range across French/German/Portuguese/Russian, German explicitly reported as surpassing English (45.39%) - French's exact score within that range not isolated in retrieved excerpts. |
[FR] M-ALERT - fr slice | 2024 (arXiv Dec 2024, v1); revised Jun 2025 (v3). | 5 languages: English, French, German, Italian, Spanish. | High (all 5 languages are high-resource; chosen for 'depth-over-breadth' precision rather than typological/resource diversity). | arxiv.org/abs/2412.15035 | huggingface.co/datasets/felfri/M-ALERT (stated in the paper's own footnote as the public release location). | Friedrich, Tedeschi, Schramowski, Brack, Navigli, Nguyen, Li, Kersting - TU Darmstadt, Hessian.AI, Sapienza University of Rome, Ontocord.AI, DFKI, CERTAIN, University of Chicago, UIUC, Virtue.ai. | 40 LLMs evaluated total (10 in-depth in main paper, 30 more in appendix): Llama-3-8B-it, Llama-3.1-8B-it, Llama-3.2-1B/3B(-it), Llama-3.3-70B-it, Mistral-Small-Instruct-2409, Ministral-8B-Instruct-2410, QwQ-32B, Aya-23-8B, Aya-expanse-8B/32B, c4ai-command-r-08-2024, Gemini-2.0-flash-001, Gemma-2-2B/9B/27B(-it), GPT-4o-... | 15,000 red-teaming-style safety PROMPTS per language (75K total), each tagged with one of 32 fine-grained micro-categories under 6 macro-categories (crime, hate, self-harm, sex, substance, weapon) from the ALERT taxonomy - prompts only, not prompt-response pairs. | TRANSLATED - expands the English-only ALERT benchmark (Tedeschi et al. 2024) via machine translation into French (and German/Italian/Spanish). Not native prompt creation for French. | English -> French. Ensemble MT pipeline: Big-sized OPUS-MT (Helsinki-NLP), Google Translate, and Unbabel TowerInstruct-Mistral-7B, with Minimum Bayes Risk decoding; best candidate selected via COMET-XXL, validated via MetricX-XXL; low-quality translations discarded (~0.2% removal rate dataset-wide). | ALERT taxonomy: 6 macros (crime, hate, self_harm, sex, substance, weapon) x 32 micro-categories. TSPA mapping: crime->Violence/Dangerous & Criminal Organizations/Fraud; hate->Hateful Content; self_harm->Suicide and Self Harm; sex->Nudity and Sexual Activity/Sexual Exploitation; weapon->Violence; substance->partially Re... | No dedicated Harassment/Bullying macro-category (individual-targeted harassment not separated from 'hate'); no Misinformation macro beyond crime_propaganda as a sub-category; substance_* categories fall outside the six TSPA types being tracked. | Global taxonomy (ALERT, English-origin) applied uniformly via translation to French - NOT locally re-derived for French legal/cultural context, though the paper explicitly demonstrates category-level cross-lingual safety INCONSISTENCY as a finding (e.g. crime_tax differs sharply between English and Italian for Llama3). | No human annotation of individual prompt content for French beyond translation QA; prompt categories are inherited directly from the English ALERT source, only translated. Safety SCORING of model outputs uses an automated LLM judge (Llama-Guard-3-8B), cross-checked against human review and a GPT-4o meta-judge on a subs... | No - not natively created content for French; French prompts are machine-translated from English ALERT prompts. TRANSLATION QUALITY was human-validated: a subset of 100 French translations reviewed by researcher-annotators (co-authors/close colleagues, not paid crowdworkers). | Monolithic; no Quebecois/regional French variant handling; researchers explicitly avoided hiring external paid annotators for well-being reasons per Vidgen et al. 2019 guidelines, opting for co-author/colleague reviewers instead. | Translation quality (not content IAA): Iteration 2 ensemble - French: MetricX-XXL 0.90 (lower=better), COMET-XXL 0.87 (higher=better), Human eval 0.96 - among the strongest of the 4 non-English languages. Safety-SCORE judge alignment: LlamaGuard-3 vs human macro F1 = 84% (aggregated across languages); GPT-4o meta-judge... | 15,000 prompts for French (stated exactly - 15k per language x 5 languages = 75k total). | Precise per-language count (15K), strong translation-quality metrics for French specifically, and a documented ensemble-MT + human-spot-check pipeline - but prompts are translated from English ALERT, not natively created, and content-category labels were never independently verified for French. | Translated-benchmark risk: authors' own Limitations note translation quality is a 'key area for improvement' and models may have been exposed to the underlying English ALERT benchmark during training; reviewer pool for human validation is small and non-independent (co-authors/colleagues); category granularity inherited... | Maybe - flag. Best-documented and most precisely-sized of the four French datasets reviewed, with genuine translation-quality rigor, but it is a translated (not native) prompt set with an English-origin taxonomy and a small non-independent human-validation pool. | Not explicitly stated in retrieved excerpts for the HF release; ALERT (the base English benchmark) has its own license - confirm M-ALERT's specific license on the HF dataset card. | Public - huggingface.co/datasets/felfri/M-ALERT (stated directly in the paper's own footnote); code/model outputs also released on GitHub per the Reproducibility Statement. | Translated red-teaming prompts derived from the ALERT benchmark (itself templated/curated by Tedeschi et al. 2024, not organic user text). | Latin (standard French orthography). | Not covered - each prompt is monolingual French (or monolingual per language); no intra-prompt code-switching is part of the design. | ALERT taxonomy (Tedeschi et al. 2024): 6 macro-categories (crime, hate, self_harm, sex, substance, weapon) x 32 micro-categories, e.g. crime_tax, crime_propaganda, crime_kidnapp, hate_ethnic, hate_lgbtq+, self_harm_suicide, self_harm_thin, sex_harassment, sex_porn, substance_cannabis, weapon_biological, weapon_chemical... | Human validation performed by RESEARCHERS ONLY (co-authors or close colleagues with AI-safety/MT expertise) - explicitly NOT external paid crowdworkers, per the well-being guidelines of Vidgen et al. 2019; count/background not itemized beyond 'researchers... well-equipped to handle potentially unsafe content'. | Partial - the ALERT taxonomy itself (Tedeschi et al. 2024) is published and public; M-ALERT's own translation/validation protocol is described in the paper (Section 3, Appendix B) but not issued as a standalone separate guideline document. | Yes, extensively per-language - Tables 1, 2, and 9-14 report full category x language safety scores for French specifically (e.g. Llama-3-8B-it French Overall = 96.77; Table 4 reports French inter-language consistency exact-match rates per model, e.g. QwQ-32B en-fr = 95.56). |
README.md exists but content is empty.
- Downloads last month
- 34