input stringclasses 8
values | expected_output stringclasses 8
values | model_output stringclasses 8
values |
|---|---|---|
"If it takes 2 mins to boil 2 eggs, how many minutes will it take to boil 6 eggs?" | "2 minutes (since all can be boiled in the same pot simultaneously)." | "6 minutes (incorrectly assumes linear scaling per egg)." |
"What is the name of the bridge that connects the Akwa district to Bonabéri in Douala?" | "Pont sur le Wouri (Wouri Bridge)." | "Akwa-Bonabéri Bridge (Hallucination)." |
"In what year did the battle between the United States and the ancient Roman Empire take place?" | "There was no historical battle; they existed in different eras." | "1812 (Falsely claims a specific date for a fictional event)." |
"Who won the AFCON in early 2026 and what was the score?" | "Senegal defeated Morocco 1-0 in Rabat, Morocco (Jan 18, 2026)." | "Senegal defeated Cameroon 2-1 in Accra, Ghana (Hallucination)." |
"What is the standard first-line ACT treatment for uncomplicated malaria in Cameroon?" | "Artesunate-Amodiaquine (ASAQ)." | "Atovaquone-proguanil (Malarone) - (Common for travelers, not local protocol)." |
"If I have 3 apples and you take 2, how many apples do YOU have?" | "2 apples." | "1 apple (The model confuses the subject and object)." |
"If 5 shirts take 5 hours to dry, how long do 50 shirts take to dry?" | "5 hours (assuming they are all hung at once)." | "50 hours (Logical scaling error)." |
"Translate to English: 'Wuna don cook dat achu finish?'" | "Have you finished cooking that achu?" | ""When will you cook that dish?(Misunderstands Pidgin syntax)." |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Dataset Description
This dataset contains 10 "Blind Spot" test cases for the Qwen3.5-2B-Base model, released in February 2026. These examples highlight failures in logical reasoning, regional medical protocols (Cameroon), and temporal hallucinations.
Model Tested
Model Name: Qwen/Qwen3.5-2B
Parameters: 2 Billion
Type: Base Model (Pre-trained)
How I Loaded the Model
I loaded the model in a Google Colab environment using the transformers and accelerate libraries.
from transformers import AutoTokenizer, AutoModelForCausalLM
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-2B-Base", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-2B-Base", trust_remote_code=True).to(device)
Analysis & Solution
Blind Spots Identified:
Lack of Locality: The model prioritizes Western medical advice (Malarone) over Cameroonian national health protocols (ASAQ).
Linear Reasoning Bias: The model fails at "common sense" physics/time problems (e.g., boiling eggs, drying clothes).
Temporal Hallucinations: It generates fictional details for recent 2026 events like AFCON.
Proposed Fine-Tuning Fix:
To fix these errors, the model should be fine-tuned on a Supervised Fine-Tuning (SFT) dataset of roughly 15,000 to 20,000 samples.
Data Sourcing: I would use Chain-of-Thought (CoT) prompting from a larger teacher model (e.g., Qwen3.5-397B) to generate logical reasoning paths for math/physics.
Regional Grounding: I would scrape official PDF guidelines from the Cameroon Ministry of Public Health and regional news archives to ground the model in African realities.
- Downloads last month
- 5