Dataset Viewer
Auto-converted to Parquet Duplicate
id
int64
1
10
input
stringclasses
10 values
expected_output
stringclasses
10 values
model_output
stringclasses
10 values
error_type
stringclasses
10 values
1
How many r's are in the word 'strawberry'? Answer with only the number.
3
The number of r's in the word "strawberry" is **2**.
letter_counting
2
Reverse this string exactly: clinical_trial_2026
6202_lairt_lacinilc
To reverse the string **"clinical_trial_2026"**, we need to reverse the characters in the string. Here's the step-by-step process: 1. The original string is: `clinical_trial_2026`. 2. Reversing the characters will give us: `62026tirnialc`. So, the reversed string is: **`62026tirnialc`**.
string_reversal
3
What day of the week was 2024-02-29? Answer with only the day.
Thursday
The day of the week corresponding to 2024-02-29 is **Tuesday**.
calendar_reasoning
4
A clinic sees 18 patients in the morning and 27 in the afternoon. If 11 are follow-ups, how many are new patients? Answer with only the number.
34
We are given: - **18 patients in the morning** - **27 patients in the afternoon** - **11 follow-ups** - **We need to find the number of new patients**. ### Step 1: Total patients seen in the day $$ 18 + 27 = 45 \text{ patients} $$ ### Step 2: Subtract the number of follow-ups from the total patients $$ 45 - ...
arithmetic
5
Translate to French: The screening result is not abnormal.
Le résultat du dépistage n’est pas anormal.
Le résultat de la surveillance n’est pas abnormal.
negation_translation
6
Which is heavier: 1 kilogram of cotton or 900 grams of iron? Answer in one short sentence.
1 kilogram of cotton is heavier.
1 kilogram of cotton is heavier than 900 grams of iron.
comparison
7
List the prime numbers between 20 and 40.
23, 29, 31, 37
The prime numbers between 20 and 40 are: **23, 29, 31, 37**.
prime_numbers
8
If all koilocytotic cells are abnormal, and this sample is koilocytotic, can it be normal? Answer with only yes or no.
No
Yes.
logic
9
Convert 98.6°F to Celsius. Answer with only the number rounded to one decimal place.
37.0
To convert 98.6°F to Celsius: $$ C = \frac{5}{9} \times (F - 32) $$ $$ C = \frac{5}{9} \times (98.6 - 32) = \frac{5}{9} \times 66.6 = 37.0 $$ **Answer:** 37.0°C (rounded to one decimal place).
unit_conversion
10
Name the capital of Australia and then give one reason many people answer incorrectly.
Canberra. Many people incorrectly say Sydney because it is more famous.
The capital of Australia is **Sydney**. Many people answer incorrectly because they might think the capital is **Melbourne** or ** Canberra**. However, **Canberra** is the capital of Australia, not the capital of any other country.
factual_precision

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

I evaluated Qwen/Qwen3-0.6B, an official Hugging Face text-generation model with 0.6B parameters. The model card lists it as a general causal language model and provides a Transformers quickstart.

The Blind spots I found

This model appeared weakest on:

exact character counting

string reversal and format preservation

calendar/date reasoning

arithmetic with multiple steps

negation-sensitive translation

constrained responses requiring exact formatting

What data could fix this?

To fix these mistakes, I would fine-tune the model on a dataset of specific instructions that includes the following:

Tasks that involve manipulating strings

reasoning with numbers and calendars

tasks that require an exact response

examples of paraphrasing or translating that are sensitive to negation

examples of minimal pairs in more than one language

How I would make the dataset

I'd combine:

synthetic examples made by a computer for counting, reversing, math, dates, and changing units

curated multilingual minimal pairs for negation and meaning preservation

examples of following instructions that are adversarial and have strict format rules

Review by hand to get rid of items that are not clear

Estimate of the size of the dataset

To see a big difference, I would start with 20,000 to 50,000 high-quality examples that are evenly spread out across these types of failures. If the goal is to make reliable generalizations instead of just a small benchmark gain, I would aim for 100,000 or more examples with held-out adversarial test sets.

Downloads last month
5