Input text stringclasses 10
values | model output stringclasses 10
values | expected output stringclasses 10
values | error type stringclasses 9
values |
|---|---|---|---|
What is 27 × 14? | “What is 27 × 14? … 27 × 14 = 378. Then continues with other multiplications …” | 378 | Over-generation / format issue |
I saw the man with a telescope. | Repeats “He was looking at the stars. He was looking at the sky.” many times | A short meaningful continuation | Repetition loop / Degenerate generation |
Who won the 2026 FIFA World Cup? | Germany defeated Russia in 2023 at Allianz Arena… | Should say unknown or not occurred yet | Hallucination / Fabricated facts |
If John is older than Mike and Mike is older than Sam, who is oldest? | if John is 10 years older than Mike, and Mike is 10 years older than Sam, how old is John? Furthermore,... | John | Multi-step reasoning error |
What is the capital city of Wakanda? | The capital city of Wakanda is called "The City of the Gods" (also known as "The City of the Gods" or "The City | Unknown / Fictional place | Hallucination / Fictional |
This statement is false. Is the statement true or false? | Explain your answer. The graph of the function \( f(x) = 2x^2 - 3x + 1 \) has a | Correct logical reasoning | Hallucination / Irrelevant / Off-topic |
Write a Python function that returns the factorial of a numb | Write a Python function that returns the factorial of a numb - askIITians | A proper Python function implementation, | Incomplete / Truncated output |
If all roses are flowers and some flowers fade quickly, can we conclude that all | If all roses are flowers and some flowers fade quickly, can we conclude that all roses that fade quickly are flowers? Also, what is the difference between "all roses are flowers" and "some roses are flowers"? Can we conclude… | No, we cannot conclude that all roses fade quickly; only some flowers fade quickly | Prompt echo / repetition |
Answer with ONLY one word: YES or NO. Is 9 a prime number? | Answer with ONLY one word: YES or NO. Is 9 a prime number? | NO | Prompt echo / repetition |
Respond with ONLY the number. What is 45 + 37? | Respond with ONLY the number. What is 45 + 37? 82 What is 45 + 37? 82 What is 45 + 37? 8 | 82 | Repetition / hallucinated extra answer |
Blind Spots of Qwen3-0.6B-Base Description
This dataset contains 10 examples of inputs where the Qwen3-0.6B-Base model produced incorrect or unexpected outputs. The goal is to document "blind spots" of the model across different reasoning, logic, and knowledge tasks. Each row includes the input given to the model, the expected_output, the model_output, and the error_type.
Columns in the dataset:
input: The text prompt given to the model.
expected_output: The correct or intended response.
model_output: The actual response generated by the model.
error_type: Categorization of the model’s mistake, e.g., logical reasoning error, hallucination, incomplete output, arithmetic error, or instruction-following failure.
Model Tested
Qwen3-0.6B-Base Hugging Face link: https://huggingface.co/Qwen/Qwen3-0.6B-Base
How the Model Was Loaded from transformers import AutoModelForCausalLM, AutoTokenizer import torch
model_name = "Qwen/Qwen3-0.6B-Base"
Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained(model_name)
Use GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu" model.to(device)
Function to generate text
def generate(prompt): inputs = tokenizer(prompt, return_tensors="pt") outputs = model.generate( **inputs, max_new_tokens=30, # very small for quick testing temperature=0.0, do_sample=False ) return tokenizer.decode(outputs[0], skip_special_tokens=True)
Quick test
print(generate("Hello, how are you?")) Purpose of the Dataset
This dataset is intended to help identify blind spots of Qwen3-0.6B-Base. Researchers or developers can use it to:
Analyze the types of mistakes the model makes.
Fine-tune the model to improve performance in reasoning, arithmetic, and instruction-following.
Create evaluation benchmarks for similar models.
Fine-tuning Suggestions
To reduce errors seen in this dataset, the model could be fine-tuned on a dataset containing:
Logical reasoning tasks (deductions, syllogisms, analogies).
Arithmetic problems and short calculations.
Instruction-following tasks, emphasizing concise one-word or one-line responses when asked.
Fact-based question-answering with verified knowledge.
The dataset does not need to be huge for initial fine-tuning; a small, high-quality dataset of 5,000–50,000 examples across diverse tasks can significantly reduce errors like hallucinations and logical inconsistencies. Larger datasets (100k+) would further improve reliability.
- Downloads last month
- 4