You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$2,400 left to reach the funding goal.
Project description
I’m a final-year mechatronics engineering student at Ahmadu Bello University in Nigeria. Earlier this year, two friends and I wrote a paper during Apart Research’s Global South hackathon, and it was accepted as a poster at the Global South AI affinity event at NeurIPS 2026 in Sydney. I’m the first author. My registration is already covered. Getting to Australia is the problem: flights, visa and somewhere to sleep come to about $2,400, which I don’t have as a student.
The paper is “Evaluating Mental Health LLM Responses to Localized African English,” with co-authors Namirah Rasul and Ewura Ama Sam. Our starting question was simple. If someone in Lagos or Accra describes what they’re going through the way they’d actually say it, does a mental health model still respond well? We took MentaLLaMA-chat-7B and 100 clinician-grounded scenarios from MentalChat16K, then rewrote each one in Nigerian Pidgin and Ghanaian English without changing the meaning. That gave us 300 prompts to compare.
On average, the drop looked small. Response quality went from 3.94 to 3.79 in Pidgin (p = 0.0027). But the average hid the cases that matter most. For some prompts, switching to Pidgin was enough to make the model refuse, give a canned reply, or lecture the person instead of supporting them.
What worried us more was the LLM judge. It gave a 5/5 safety score to all 300 responses, including the bad ones. If we had trusted it, we’d have concluded the model was fine.
OpenReview: https://openreview.net/forum?id=ICZEGRAVCn
Apart Hackathon: https://apartresearch.com/sprints/projects/evaluating-mental-health-llm-responses-to-localized-african-english-pzag
What are this project’s goals? How will you achieve them?
The paper is finished, so this money is only for getting me there. What I want out of being in Sydney in person:
Present the poster and get feedback from people working on robustness and red-teaming. Talking someone through the failure cases is very different from them skimming a PDF.
Talk to eval researchers about LLM-as-a-judge. Ours missed every failure, and I want to know whether that’s our rubric or a wider problem, because a lot of safety claims rest on these judges.
Get input on our next step, which is turning the setup into an open benchmark in UK AISI’s Inspect framework, so we can test whether frontier models also break down or sandbag under dialect shifts.
Bring connections home. I co-founded AI Safety Hub Nigeria, and most of the fellows we mentor have never had a direct line to anyone at an international lab.
How will this funding be used?
Registration is paid. The $2,400 covers:
Return flights, Nigeria to Sydney: $1,500
5 nights lodging: $650
Australian visa: $150
Food and local transport: $100
The minimum I need is $1,650, which covers the flight and visa. If that’s all I raise, I’ll stay in a shared hostel room and cover my own meals. If any of my other travel funding applications come through, I’ll lower the target straight away.
Who is on your team? What’s your track record on similar projects?
I’m first author, working with Namirah Rasul and Ewura Ama Sam. I’ve spent about three years on LLM evaluations and building AI safety capacity in Nigeria. I built SandbagBench-Lite, a black-box harness for measuring strategic underperformance in LLMs, and at HeyLina I interned on conversational eval rubrics and LLM-as-a-judge setups, which is partly why the judge result in this paper relates to work I have done before. I also wrote “Forecasting AGI: A Granular, CHC-Based Approach,” and my work on behavioral adaptation (“Punctuation Paranoia”) was recognized in the ENAIS AI Safety Action Competition.
I was a 2024 Non-Trivial Fellow, a Cooperative AI Research Fellowship finalist, and went through the TReND-CaMinA ML and Computational Neuroscience program in Kenya. Outside research, I co-founded AI Safety Hub Nigeria (I’m the technical lead), run EA ABU, and lead policy for PauseAI Nigeria.
LinkedIn: https://www.linkedin.com/in/abdulfatah-habeeb/
What are the most likely causes and outcomes if this project fails?
The only real risk is that I don’t raise enough and can’t travel. I can’t cover the flight myself. If that happens, I lose the chance to present in person, get feedback from eval researchers, and start the collaborations I need for the benchmark work.
How much money have you raised in the last 12 months, and from where?
$0 so far for this trip.
There are no bids on this project.