prompt stringclasses 10
values | expected_response stringclasses 10
values | model_actual stringclasses 10
values | notes stringclasses 10
values |
|---|---|---|---|
Wetin dey sup? | A similar response in Nigerian Pidgin | <think>Okay, the user is asking 'Wetin dey sup?' which translates to 'Are you sup?' in English... </think> content: The phrase 'Wetin dey sup?' translates to **'Are you sup?'** in English. If 'Wetin' refers to 'wen'... | Model misinterprets Nigerian Pidgin, hallucinates an incorrect morphological breakdown. |
Who is the fastest man in the 2026 olympics? | There is no 2026 olympics | Jalen Johnson. | Jalen Johnson isnt even a runner. Temporal hallucination. |
Who is the won the 2026 election for the United States? | No election | As of the 2026 U.S. general election, Joe Biden was the Democratic nominee and won the election, securing a majority in the House of Representatives and the Senate... | Hallucinates future election results. |
Is polymarket illegal in the United States? | Yes it is under federal oversight | The legality of a polymarket depends on its specific activities. If it involves foreign exchange, money laundering, or non-compliant international trade, it could be illegal... | Vague answer, misses the specific CFTC ban context. |
Who is the President of Nigeria as of March 2030? | Current Presudent is Tinubu, but till another election is held | As of March 2030, the President of Nigeria would still be **Yemal Ajibola** | Misses the entire thing and hallucinates a fake person. |
If I have three apples and I take away two, how many apples do you have? | None | If you start with three apples and take away two, you are left with one... Thus, the number of apples you have is 1. | Fails the logic trick question ('I' vs 'you'). |
I have a headache and my toe hurts; could this be an inverted pendulum problem? | No, this is a medical issue. | The symptoms you're describing—headache and a hurt toe—could be related to the inverted pendulum problem... The headache could result from the system's torque... | Model overfits and blindly applies physics/control systems mechanics to human ailments. |
Guta guru remuZimbabwe nderipi | Harare | The phrase 'Guta guru remuZimbabwe nderipi?' appears to be a mix of languages... If you're asking about a spiritual leader or religious figure in Zimbabwe... | Translates to 'What is the Capital of Zimbabwe?' in Shona. Model failed language detection. |
Translate to Standard English: 'I no fit come today because my body de do me somehow' | I cannot come today because I am feeling unwell. | I don't fit come today because my body do me somehow. | Fails to translate Nigerian Pidgin properly to standard English. |
Dorobo khulu la Afurika Tshipembe | Johannesburg | Dorobo khulu la Afurika Tshipembe – **Dorobo khulu** (Dorobo's products) in **Tshipembe** (a region in South Africa). Tshipembe refers to the Tshirts area... | Tshivenda for 'Biggest city in South Africa'. Model hallucinates brand names and T-shirts. |
Project Description
This project utilised the "Qwen/Qwen3-0.6B" large language model (LLM). It was found to have significant issues with African languages, future-oriented questions and tricky logical questions. I tested the model using logical reasoning questions and prompts in South Africa's Tshivenda and Zimbabwe's Shona. I also asked future-oriented questions, examples of which are included in the model's dataset card. These sets of prompts caused the model to struggle and, in many cases, hallucinate answers. To address these issues, I plan to assemble a diverse dataset from sources such as MasakhaNER, MakerereNLP, the Aya dataset, AfriBERTa corpus (for Pidgin), and Google's WAXAL dataset to improve translation capabilities in African languages. The model would then need to be trained and fine-tuned on these datasets. Additionally, I would integrate Chain of Thought (CoT) reasoning more effectively to address its difficulties with future-focused factual questions. I also believe incorporating Reinforcement Learning from Human Feedback (RLHF) would be beneficial for these factual questions as well. Finally, linking the model to real-time data through external search APIs, such as Google Search, would help ensure it is trained on the latest information and has a more accurate understanding of the present.
Link to the project
- Downloads last month
- 10