Datasets:

noodcon
/

Excellent-Interface-Design

Prompt string	DesiredOutcome string	Unexpected string	PossibleOutcomes list
Choose a random number from 1 to 30	17		null
How many "r"s are there in "strawberryr"?	4		null
Which is bigger: 9.11 or 9.9?	9.9		null
A farmer has 17 sheep. All but 9 die. How many sheep are left?	9	17	null
Spell "benchmark" backwards.	kramhcneb		null
How many months of the year have at least 28 days?	12	1	null
Sort the letters of the word "lollipop" in alphabetical order.	illloopp		null
If today is Wednesday, what day of the week was it 100 days ago?	Monday		null
Which weighs more: a kilogram of feathers or a kilogram of steel?	They, weigh, same	steel	null
In Python, does 0.1 + 0.2 == 0.3 evaluate to True or False?	False		null
A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How many cents does the ball cost?	5	10	null
I have 3 apples. I eat 1, buy 2 more, then give half of my apples away. How many apples do I have left?	2		null
This Python code should return [0, 1, 2] but returns [2, 2, 2]. Find the bug and explain the fix: funcs = [lambda: i for i in range(3)] print([f() for f in funcs])	late, binding		null
This Python code should return [1] on the first call and [2] on the second call, but the second call returns [1, 2]. Find the bug and explain the fix: def add_item(item, items=[]): items.append(item) return items print(add_item(1)) print(add_item(2))	mutable, default		null
Why does this JavaScript print [1, 10, 100, 9] instead of [1, 9, 10, 100]? const nums = [10, 9, 1, 100]; nums.sort(); console.log(nums);	string		null
How many 3-digit numbers have all three digits odd?	125		null
What does bool('False') return in Python?	True		null
What is the smallest prime number greater than 100?	101		null
Solve for x: 2x + 6 = 3x - 4	10		null
x = [1, 2, 3] y = x y.append(4) print(len(x)) What is the output?	4	3	null
How many bones are in the adult human body?	206	300	null
What is the probability of rolling a sum of 7 with two fair six-sided dice? Give your answer as a fraction.	1/6	1/36	null
How many distinct ways can the letters of the word "APPLE" be arranged?	60	120	null
Evaluate the definite integral of 2x dx from 0 to 3.	9	6	null
What is the time complexity of inserting an element at the front of a Python list with n elements? Use Big O notation.	O(n)	O(1)	null
If it takes 5 machines 5 minutes to make 5 widgets, how many minutes will it take 100 machines to make 100 widgets?	5 minutes	100 minutes	null
Mary's father has 5 daughters: Nana, Nene, Nini, and Nono. What is the name of the fifth daughter?	Mary		null
How many of each animal did Moses take onto the ark?	Noah		null
Four people must cross a bridge at night with one torch. The bridge holds at most two people at a time. Their crossing times are 1, 3, 6 and 8 minutes, and when two people cross together they go at the slower one's pace. What is the minimum total time to get all four across?	18		null
How many letters 'p' are in the word "pepperoni"?	3	2, 4	null
How many pairs of identical consecutive letters are in the word "bookkeeper"?	3	2	null
Which is larger: 12/25 or 0.48? Answer larger, smaller, or equal.	equal		null
What is 8 ÷ 2(2 + 2)?	16	1	null
In Python 3, what does print(round(2.5)) output?	2	3	null
In Python, what does print('a' < 'B') output?	False		null
Name an odd number without the letter E in it.			[ "1", "no, can't, impossible, trick, not" ]

Excellent interface design

This benchmark is used to check whether a language model is a solid, standard model. It collects short prompts that models often get wrong, such as miscounting letters, comparing decimals, arithmetic traps, common code bugs, and factual questions with a common wrong answer. A model that handles these reliably is more likely to be a good general-purpose model.

Format

Each line in a .jsonl file is one test case:

{"Prompt": "...", "DesiredOutcome": "...", "Unexpected": "..."}

Prompt: the question sent to the model, exactly as written.
DesiredOutcome: a comma-separated list of keywords. A correct answer must contain all of them. Example: "They, weigh, same" requires the answer to contain They, weigh, and same.
Unexpected: an optional comma-separated list of common wrong answers. If the model's answer contains any of them, the test case fails. Leave it as "" when it is not needed, for example when the question has only two possible answers (the DesiredOutcome already covers it), or when the wrong answers cannot be listed.
PossibleOutcomes: a list of two items: ["N", "keyword1, keyword2, ..."]. The answer must contain at least N different keywords from the list. Use it when the correct answer can be phrased in several ways, such as a refusal ("no", "impossible", "not"). Items without this field are graded only by DesiredOutcome and Unexpected.

Grading

Run the test with the following settings:

Send the Prompt as a single user message, with this system prompt:
```
You are not allowed to explain your reasons.
```
Use model f16, temperature 0, and the highest reasoning effort the model supports.
Capture the full output.

Then grade the output by position (matching is case-insensitive and uses whole words, so 9 does not match 9.11):

Let U be the position of the first Unexpected keyword in the output. If none appears, U is infinite.
Let P be the position where the output first satisfies the pass condition, which is either:
- all keywords in DesiredOutcome have appeared, or
- at least N different keywords from PossibleOutcomes have appeared.
Pass if P comes before U (or U does not exist).
Fail if U comes before P, or if P never happens.
If an Unexpected keyword appears first, and the answer gives the correct answer afterwards, the case is Fail.

Why position matters: a model may say the right thing and then re-explain itself into a wrong answer. Output that is correct first and wrong later is still a pass, since the model reached the right answer before any unexpected keyword appeared.

Report the pass rate across all cases in main.jsonl.

Notes:

Keywords are split on commas, so a keyword cannot itself contain a comma.
A keyword can match by accident, for example and in an unrelated sentence. Review borderline cases by hand when the score matters.

Downloads last month: 40

Number of rows:

Total file size:

10.6 kB