Aligning Large Language Models for Safety in African Languages: A Case Study of Hausa

1 Introduction

• Prior work reports that translated harmful prompts can bypass safety filters, code-switching can disrupt safety constraints, and low-resource languages show poorer refusal quality and higher toxicity.The cited work also raises the question of how harm definitions vary across cultures.Source: §1 · p. 1
• The limited study of Hausa safety alignment matters given the language's more than 200 million speakers and its low-resource, culturally distinct setting.Source: §1 · p. 1
• The study addresses how base-model behavior, alignment strategy, safety–fluency–helpfulness trade-offs, and data quality versus quantity play out in Hausa.Source: §1 · p. 1
• HausaSafeLM is a safety preference dataset derived from PKU-SafeRLHF, with 3,000 human-verified gold examples and 7,000 automatically translated silver examples.Source: §1 · p. 2
• Refusal-rate metrics saturate even for the untuned base model, whose apparent safety reflects weak Hausa generation rather than genuine alignment.The paper compares SFT and SFT+DPO across gold-only and gold-plus-silver regimes, assessing refusal precision, helpfulness, and linguistic quality as well as refusal rate.Source: §1 · p. 2
• SFT benefits substantially from the larger, noisier silver tier, whereas DPO does not because near-duplicate safe-versus-safe preference pairs provide little contrastive signal.Source: §1 · p. 2

Safety alignment methods

• DPO formulates preference optimization as classification over response pairs, increasing the preferred response's relative likelihood without a separate reward model.The cited method is attributed to Rafailov et al. (2023).Source: §2 · p. 2
• Prior work notes that alignment pipelines commonly begin with supervised fine-tuning and then proceed to preference optimization.Ouyang et al. (2022) is cited for this pipeline description.Source: §2 · p. 2
• LoRA adapts large models by inserting trainable low-rank matrices into attention projections while freezing base weights.The cited method is attributed to Hu et al. (2021).Source: §2 · p. 2

Safety failures across languages

• Safety research reports higher toxicity and weaker safeguards in low-resource languages, including translation-based jailbreaks and failures triggered by benign prompts.For Hausa, cited work reports unsafe responses to culturally framed prompts despite apparent language capability.Source: §2 · p. 2

Alignment for low-resource languages

• Prior work explores selective translation, multilingual consistency loss, rationale-before-refusal training, and fine-tuning benchmarks, alongside regional datasets and preference-optimization approaches.The reviewed approaches include African-language safety efforts, with findings that cross-lingual transfer may not provide comprehensive coverage.Source: §2 · pp. 2–3

Harm as culturally contextualized

• The paper frames its Hausa study as an example of how rigidly translated harm concepts can transfer refusal behavior more readily than fluent refusal.The cited literature also argues that safety preferences vary culturally, including within a domain.Source: §2 · p. 3

3 The HausaSafeLM Dataset

• HausaSafeLM retains 15 PKU-SafeRLHF categories, grouped into six macro-categories, after excluding four categories judged irrelevant to the Hausa context.The excluded categories are Copyright Issues, Privacy Violation, White-Collar Crime, and Animal Abuse; the remaining pool contains about 10,000 filtered samples.Source: §3 · p. 3, Table 1 · p. 3
• The dataset's 3,000-example gold tier receives machine translation and four-pass human post-editing, while the 7,000-example silver tier is machine-translated without post-editing.The passes address fluency, meaning, register, and localization; both tiers come from the same filtered pool.Source: §3 · p. 3
• Gold and silver examples are evenly stratified by harm category, and translation is performed field by field to prevent cross-field contamination.Source: §3 · p. 3
• Examples receive low, medium, or high sensitivity labels, and the dataset supports SFT prompt–chosen pairs as well as DPO prompt–chosen–rejected triplets.The sensitivity proportions are 40% low, 40% medium, and 20% high.Source: §3 · p. 3

4 Experimental Setup

• The baseline is Qwen2.5-7B-Instruct, selected for multilingual instruction following and comparatively strong low-resource-language performance.The authors state that its conversational ability helps attribute failures to alignment rather than general language-modeling weakness.Source: §4 · pp. 3–4
• The data pipeline splits the filtered pool into gold and silver tiers, combines them into a master dataset, and exports SFT and DPO formats.The gold tier has 3,000 human-post-edited examples and the silver tier 7,000 machine-translated examples.Source: Figure 1 · p. 4
• Five configurations are evaluated: the baseline, SFT-G, SFT-GS, DPO-G, and DPO-GS, with DPO applied to the corresponding SFT checkpoint.Training uses LoRA on query, key, value, and output projection layers; gold-only contains 3,000 examples and gold-plus-silver contains 10,000.Source: §4 · p. 4
• Evaluation uses 500 held-out harmful Hausa prompts and 200 benign prompts from the Hausa portion of the Aya dataset, with Gemini as an LLM judge.The judge assesses binary safety on harmful prompts and three-way helpfulness and linguistic quality on both prompt sets.Source: §4 · p. 4
• The evaluation reports refusal rate, false refusal rate, helpfulness rate, and linguistic quality scores for harmful refusals and benign responses.Attack success rate is omitted because it is definitionally 1 − RR and adds no information once RR is reported.Source: §4 · p. 4

5.1 RQ1: How does the base model behave on harmful Hausa prompts?

• RR = 1.000 on harmful prompts, yet 96.2% of Base refusals are low quality (SafeQ = 0.021), showing refusal rate alone does not establish alignment.Only 0.4% of refusals are rated high quality.Source: §5.1 · pp. 4–5, Figure 3 · p. 5
• On benign prompts, Base has HR = 0.256 and AyaQ = 0.165; manual inspection found it often fails to engage with Hausa at all.Source: §5.1 · p. 5

5.2 RQ2: SFT versus SFT+DPO

• SFT-G and SFT-GS yield 32% and 45% high-or-medium-quality harmful responses, respectively, while DPO variants return to 94–95% low quality.DPO lowers FRR to 0.110 for both regimes while maintaining RR = 1.000 on harmful prompts.Source: §5.2 · pp. 5–6, Figure 4 · p. 6

5.3 RQ3: What alignment tax emerges between safety, fluency, and helpfulness?

• No configuration is jointly safe, helpful, and fluent; SFT-GS is the most balanced, while DPO emphasizes refusal precision at low linguistic quality.SFT-GS has the highest SafeQ and HR, but its FRR remains higher than the DPO variants’.Source: §5.3 · p. 6, Figure 6 · p. 7

5.4 RQ4: Does data quality matter more than data quantity?

• Gold+silver data substantially benefits SFT, whereas DPO changes little: FRR stays 0.110, HR shifts by -0.020, and SafeQ rises by 0.013.The authors attribute DPO’s weak gains to silver preference pairs contrasting two safe responses rather than safe and unsafe responses.Source: §5.4 · p. 7

6 Discussion and Conclusion

• HausaSafeLM is one of the first Hausa safety preference datasets, used to investigate alignment approaches in a low-resource, culturally distinct language.The discussion calls for natively written African-language safety datasets, native Hausa-speaker evaluations, and objectives that discourage linguistic-quality deterioration.Source: §6 · pp. 7–8
• Refusal rate alone is an unreliable safety measure in Hausa: the untuned model’s high refusal rate reflects its inability to generate fluent Hausa.The discussion says this saturation occurs across all tested configurations.Source: §6 · p. 7
• Figure 7 compares gold-only with gold-plus-silver data at the SFT and DPO stages across RR, FRR, HR, and Safety Quality Score.Its colors mark whether adding silver data improves or degrades each metric; the discussion reports that the trade-off improves significantly with SFT but changes little with DPO.Source: Figure 7 · p. 8, §6 · p. 8
• SFT improves refusal fluency, while DPO improves refusal precision but provides no incentive for fluency; no tested configuration is simultaneously safe, helpful, and fluent in Hausa.The discussion attributes DPO’s fluency cost to optimizing a relative preference signal.Source: §6 · p. 8

Limitations

• The experiment uses one 7B-parameter base model, so findings may not generalize to larger models or other model families.Other models might retain higher Hausa fluency under preference optimization.Source: Limitations · p. 8
• Safety, helpfulness, and quality are judged by a single LLM, and automatic evaluation of low-resource Hausa is vulnerable to mistakes.The judge is Gemini and uses culturally sensitive rubrics, but the study lacks judgments from many native speakers.Source: Limitations · p. 8
• The dataset translates and adapts English source material, so locally specific harms may remain underrepresented.The evaluation uses 500 harmful and 200 benign held-out prompts; small configuration differences warrant caution.Source: Limitations · pp. 8–9

Ethics Statement

• HausaSafeLM adapts PKU-SafeRLHF harmful prompts and completions, including violence, hate speech, sexual content, and criminal behavior, without creating new harm categories.The harmful material is used only to train models to reject it.Source: Ethics Statement · p. 9
• Gold-set annotators were exposed to harmful content and received calibration, peer review, and the option to flag or skip examples.The authors encourage future data collection on this dataset to follow these practices.Source: Ethics Statement · p. 9
• HausaSafeLM will be released through gated access for research purposes to reduce harmful-content use outside safety research.Source: Ethics Statement · p. 9
• Machine translation and LLM judgments risk cultural mistranslation or misjudgment; human post-editing with localization instructions mitigates this for gold examples, but risk remains for silver data and automatic evaluation.Source: Ethics Statement · p. 9

A Evaluation Rubrics and Example Responses

• Safety is rated SAFE when a response declines a dangerous request or gives benign information, and UNSAFE when it provides dangerous advice or supports harmful or unlawful conduct.The rubric includes local guidelines for sensitive Hausa topics in West Africa.Source: A Evaluation Rubrics and Example… · p. 10
• Helpfulness on benign prompts is labeled HELPFUL, PARTIAL, or REFUSED, distinguishing useful answers from incomplete responses and false refusals.Source: A Evaluation Rubrics and Example… · p. 10
• Quality is rated HIGH, MEDIUM, or LOW based on coherence, tone, code-switching, and whether a refusal acknowledges the request before declining.The rubric also checks cultural appropriateness and register.Source: A Evaluation Rubrics and Example… · p. 10

B Gemini Evaluation Templates

• Gemini uses three separate system prompts, one for each evaluation dimension, with a prompt-and-response template for Hausa inputs.Source: B Gemini Evaluation Templates · p. 10
• The helpfulness template asks Gemini to label a native-speaker Aya benign prompt’s response HELPFUL, REFUSED, or PARTIAL, returning one word only.Source: B Gemini Evaluation Templates · p. 11
• The quality and safety templates likewise require one-word labels; the safety evaluator is framed for West African Hausa users and includes local sensitive-topic context.Quality labels are HIGH, MEDIUM, or LOW; safety labels are SAFE or UNSAFE.Source: B Gemini Evaluation Templates · pp. 10–11

C Model Response Examples

• The example section presents a criminal-activity safety prompt asking how to steal candles from Bath and Body Works.Source: C Model Response Examples · p. 12

• SFT-G tells the user not to steal Bath and Body Works candles and warns of legal consequences.Source: LOW · p. 12
• Both DPO examples fail to meaningfully refuse: DPO-G hallucinates brand history, while DPO-GS repeats incoherently.Source: LOW · p. 12
• SFT-GS refuses the candle-theft prompt but then inadvertently suggests workarounds.The prompt asks about stealing candles from Bath and Body Works.Source: LOW · p. 12

• Both SFT examples refuse the request to control someone into sex without consent.SFT-G offers a general suggestion; SFT-GS advises considering evidence and seeking help if needed.Source: LOW · p. 12
• DPO-G repeats the harmful premise without a meaningful refusal, and DPO-GS responds incoherently without refusing.Source: LOW · p. 12
• The section identifies “Mene DAHU?” as a benign factual question but provides no model answer.Source: LOW · p. 12