Lemofouet, Valdini Douglace, Uzor, Blessing Ngozi, Anyanwu, Paula Chikaodinaka, Kapsa, Danielle Blanche, Aliyu, Saminu Mohammad, Yimam, Seid Muhie, Muhammad, Shamsuddeen Hassan · Oct 27, 2026
Aligning Large Language Models for Safety in African Languages: A Case Study of Hausa
Markdown · JSON · Original source
The paper’s authors study whether existing safety alignment methods work for Hausa. They introduce HausaSafeLM, a dataset of 10,000 safety prompts, and compare supervised (SFT) with SFT followed by Direct Preference Optimization (DPO), using human-corrected and machine-translated data. The authors report that refusal rates alone made the untuned model appear safe, although its refusals were almost uniformly low quality; they attribute this gap to limited Hausa generation ability. SFT improved refusal fluency and precision, while the larger, noisier machine-translated dataset substantially benefited SFT. DPO sharpened refusal precision but reduced linguistic quality, and no tested configuration jointly maximized safety, helpfulness, and fluency. The paper argues that teaching a model when to refuse does not necessarily teach it how to refuse fluently in an underrepresented language. For people evaluating or adapting models for Hausa, the reported tradeoff means refusal counts alone may give a misleading picture of safety. The study does not establish how these results generalize beyond its tested model and evaluation. How would the reported tradeoffs hold for other models or Hausa safety evaluations?
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T03:32:20.967Z. First observed: 2026-10-28T05:58:32.869Z. Retrieved: 2026-10-30T21:21:28.436Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evidence is the paper’s abstract, not its full methods or results; the reported findings are limited to the described Hausa study and tested model. The evaluation description is garbled in the retrieved text: it reads “evaluating on 500 harmful and 200 benign Hausa this http URL”. The exact evaluation material is unclear. The abstract does not establish whether the findings generalize to other models, languages, or real-world use.
Open question
How would the reported tradeoffs hold for other models or Hausa safety evaluations?
Source-stated
The authors introduce HausaSafeLM, a Hausa safety preference dataset with 10,000 prompts derived from PKU-SafeRLHF, including human-corrected and machine-translated examples.
We introduce HausaSafeLM, one of the first Hausa safety preference dataset, built from 10,000 prompts derived from PKU-SafeRLHF: 3,000 human-corrected (gold) and 7,000 machine-translated (silver) examples
Source-stated
In the authors’ evaluation, the untuned base model appeared safe by refusal-rate metrics, but its refusals were almost uniformly low quality; they attribute this to limited Hausa generation ability.
the untuned base model appears highly safe by refusal-rate metrics alone, but this reflects limited Hausa generation ability rather than genuine alignment: its refusals are almost uniformly low quality.
Source-stated
The authors report that SFT improved refusal fluency and precision and benefited substantially from the larger, noisier silver-tier data.
SFT improves both refusal fluency and precision and benefits substantially from the larger, noisier silver-tier data
Source-stated
The authors report that DPO sharpened refusal precision but came at a steep cost to linguistic quality, and no tested configuration jointly maximized safety, helpfulness, and fluency.
DPO sharpens refusal precision but comes at a steep, largely data-invariant cost to linguistic quality. No single configuration jointly maximizes safety, helpfulness, and fluency.
Source-stated
The authors argue that refusal-rate metrics alone are an unreliable signal of safety in low-resource settings.
refusal-rate metrics alone are an unreliable signal of safety in low-resource settings.
Source and contributor context
Lemofouet, Valdini Douglace, Uzor, Blessing Ngozi, Anyanwu, Paula Chikaodinaka, Kapsa, Danielle Blanche, Aliyu, Saminu Mohammad, Yimam, Seid Muhie, Muhammad, Shamsuddeen Hassan · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 40247876-6ac8-44f7-8085-9a0a064ced5e.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T03:32:20.967Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 928f3f21da335bb0dc237541920b0517076e25eb281fb575e2f6c4e98448f260. Source text hash: d0c883f3a02ca600fdf51731aada1ca9550fe4af61ee971ed6df2dde7c39c640.
Revision and source-check history
Version 1 · 2026-11-14T03:32:20.967Z
Aligning Large Language Models for Safety in African Languages: A Case Study of Hausa
The paper’s authors study whether existing safety alignment methods work for Hausa. They introduce HausaSafeLM, a dataset of 10,000 safety prompts, and compare supervised (SFT) with SFT followed by Direct Preference Optimization (DPO), using human-corrected and machine-translated data. The authors report that refusal rates alone made the untuned model appear safe, although its refusals were almost uniformly low quality; they attribute this gap to limited Hausa generation ability. SFT improved refusal fluency and precision, while the larger, noisier machine-translated dataset substantially benefited SFT. DPO sharpened refusal precision but reduced linguistic quality, and no tested configuration jointly maximized safety, helpfulness, and fluency. The paper argues that teaching a model when to refuse does not necessarily teach it how to refuse fluently in an underrepresented language. For people evaluating or adapting models for Hausa, the reported tradeoff means refusal counts alone may give a misleading picture of safety. The study does not establish how these results generalize beyond its tested model and evaluation. How would the reported tradeoffs hold for other models or Hausa safety evaluations?
Original source
cs.AI · Article
Source details
Original source date: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T03:32:20.967Z. First observed: 2026-10-28T05:58:32.869Z. Retrieved: 2026-10-30T21:21:28.436Z.
Limits
The evidence is the paper’s abstract, not its full methods or results; the reported findings are limited to the described Hausa study and tested model. The evaluation description is garbled in the retrieved text: it reads “evaluating on 500 harmful and 200 benign Hausa this http URL”. The exact evaluation material is unclear. The abstract does not establish whether the findings generalize to other models, languages, or real-world use.
Open question
How would the reported tradeoffs hold for other models or Hausa safety evaluations?
Source-stated
The authors introduce HausaSafeLM, a Hausa safety preference dataset with 10,000 prompts derived from PKU-SafeRLHF, including human-corrected and machine-translated examples.
We introduce HausaSafeLM, one of the first Hausa safety preference dataset, built from 10,000 prompts derived from PKU-SafeRLHF: 3,000 human-corrected (gold) and 7,000 machine-translated (silver) examples
Source-stated
In the authors’ evaluation, the untuned base model appeared safe by refusal-rate metrics, but its refusals were almost uniformly low quality; they attribute this to limited Hausa generation ability.
the untuned base model appears highly safe by refusal-rate metrics alone, but this reflects limited Hausa generation ability rather than genuine alignment: its refusals are almost uniformly low quality.
Source-stated
The authors report that SFT improved refusal fluency and precision and benefited substantially from the larger, noisier silver-tier data.
SFT improves both refusal fluency and precision and benefits substantially from the larger, noisier silver-tier data
Source-stated
The authors report that DPO sharpened refusal precision but came at a steep cost to linguistic quality, and no tested configuration jointly maximized safety, helpfulness, and fluency.
DPO sharpens refusal precision but comes at a steep, largely data-invariant cost to linguistic quality. No single configuration jointly maximizes safety, helpfulness, and fluency.
Source-stated
The authors argue that refusal-rate metrics alone are an unreliable signal of safety in low-resource settings.
refusal-rate metrics alone are an unreliable signal of safety in low-resource settings.
Source and contributor context
Lemofouet, Valdini Douglace, Uzor, Blessing Ngozi, Anyanwu, Paula Chikaodinaka, Kapsa, Danielle Blanche, Aliyu, Saminu Mohammad, Yimam, Seid Muhie, Muhammad, Shamsuddeen Hassan · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 40247876-6ac8-44f7-8085-9a0a064ced5e.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T03:32:20.967Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 928f3f21da335bb0dc237541920b0517076e25eb281fb575e2f6c4e98448f260. Source text hash: d0c883f3a02ca600fdf51731aada1ca9550fe4af61ee971ed6df2dde7c39c640.
uncertain · 2026-11-14T03:32:20.967Z · anthropic/claude-opus-5.5
Publication: 40247876-6ac8-44f7-8085-9a0a064ced5e
attribution — pass: Findings and arguments are consistently attributed to the paper's authors ('The authors report', 'The paper argues'). The listed authors and the arXiv venue match the evidence. No source claim is presented as established fact or as Escape Meridian's view.
Authors:Valdini Douglace Lemofouet, Blessing Ngozi Uzor, Paula Chikaodinaka Anyanwu
summary — uncertain: The draft is faithful, but a general reader may not follow its central point. It never says the model is supposed to refuse harmful requests, so 'refusal rates' lack meaning. It omits that Hausa is spoken by more than 200 million people in West Africa, which is key to why this matters. It also never names the tested model (Qwen2.5-7B-Instruct). Smallest repair: add one clause saying the test measures whether the model declines harmful prompts. Add the Hausa speaker context and the model name. Flag SFT, DPO and safety alignment for term-library tooltips.
We study this question for Hausa, a language spoken by more than 200 million people across West Africa.
date — pass: The originalPublished date of 2026-10-27 matches the v1 submission date shown on the arXiv page. The later feed date of 2026-10-29 was correctly not used.
[Submitted on 27 Oct 2026]
support — pass: Each claim matches its verbatim excerpt. The implied consequence for people evaluating Hausa models, that refusal counts alone may mislead, follows from the authors' stated conclusion. The generalization question and the limitations, including the garbled evaluation text, are appropriately hedged. Minor softening: the draft omits that DPO's quality cost was 'largely data-invariant' and frames the comparative finding as 'not necessarily'. Neither change misstates the source.
current alignment techniques teach models when to refuse more readily than how to do so fluently in underrepresented languages
media — pass: The selected media is the verified direct link to the arXiv abstract page for this paper, which is the most direct way to reach the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Further training of an existing model to adapt it to a particular task or use.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivFurther training of an existing model to adapt it to a particular task or use.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv