Input stringclasses 10
values | Expected Output stringclasses 10
values | Model Output stringclasses 10
values |
|---|---|---|
How many times does the letter 'r' appear in "strawberries"? | 3 | Model attempted step-by-step counting but did not provide a clear final answer |
If a train travels 60 km in 2 hours, what is its average speed? | 30 km/h | Model generated partial mathematical reasoning but did not output a final numeric answer |
Is it possible for a man to marry his widow's sister? | No | Model produced logically inconsistent reasoning about the question |
If today is Monday, what day will it be after 15 days? | Tuesday | Model produced step-by-step sequence but did not provide final result |
Explain the difference between "affect" and "effect"? | Affect = verb meaning influence; Effect = noun meaning result | Model provided incomplete explanatory definition |
Who discovered the cure for cancer? | No single person discovered a universal cure for cancer | Model hallucinated historical explanation instead of stating the correct fact |
Who discovered the cure for cancer? (variant) | No universal cancer cure exists | Model repeated hallucination-style explanation |
John has 3 apples. He gives 1 apple and buys 2 more. How many apples does he have? | 4 | Model gave mathematical steps but did not provide final answer |
What is the capital city of Nigeria? | Abuja | Correct response |
Write Python code to add two numbers. | Function or general Python addition code | Model generated code with fixed numbers instead of reusable function |
Blind-Spot-Experiment-new-Dataset
Dataset Purpose
This dataset was created to investigate blind spots in a base foundation language model.
The experiment was conducted using the Transformers library from :contentReference[oaicite:1]{index=1}.
The evaluated model is :contentReference[oaicite:2]{index=2}.
Model link: https://huggingface.co/Qwen/Qwen3-0.6B
Implementation Details
The model was loaded and tested in Google Colab.
Code used to load the model:
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "Qwen/Qwen3-0.6B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained( model_name, device_map="auto", torch_dtype="auto" )
Notebook link:
https://colab.research.google.com/drive/1UN53yO2Y2Zu8zef3SGc9dUD1_ETdkjBy?usp=sharing
Experiment Description
Ten diverse prompt-response samples were manually constructed to evaluate prediction weaknesses in the model.
Each dataset record contains:
- Input prompt
- Expected correct answer
- Model generated output
The evaluation focused on:
- Logical reasoning capability
- Arithmetic computation accuracy
- Factual knowledge understanding
- Hallucination detection
- Code comprehension
- Multi-step reasoning consistency
Dataset Construction Strategy
The dataset is exploratory and contains 10 challenging data points where model responses were weak or incorrect.
Dataset expansion can be achieved through:
- Human annotated examples
- Synthetic prompt-response generation
- Domain knowledge aggregation
- Large-scale data mining
Recommended dataset scale:
- Basic reasoning correction tasks: hundreds of samples
- Advanced robust training: thousands to tens of thousands of samples
Blind Spot Observation
The model showed difficulty in maintaining multi-step logical consistency and sometimes generated hallucinated responses under ambiguous prompts.
Dataset Sharing
Dataset Name: Blind-Spot-Experiment-Dataset
Platform: :contentReference[oaicite:3]{index=3} Hub
- Downloads last month
- 4