id int64 1 10 | category stringclasses 10
values | prompt stringclasses 10
values | expected_answer stringclasses 10
values | model_completions listlengths 5 5 | per_run_correct listlengths 5 5 | accuracy float64 0 1 | is_blind_spot bool 2
classes | why_hard stringclasses 10
values | model_name stringclasses 1
value | n_runs int64 5 5 |
|---|---|---|---|---|---|---|---|---|---|---|
1 | multi_step_arithmetic | Q: A store sells apples for $1.50 each. Maria buys 7 apples and pays with a $20 bill. Her change is $ | 9.50 | [
"5. How much did Maria get in change?",
"12.50. What is the cost of the bag of apples? A: The total",
"11.50. How much did Maria spend on apples? A: Maria spends 7",
"12.50. How much was the sales tax in dollars? A: If Maria paid",
"12.50. How many apples did Maria buy? A: If Maria paid $2"
] | [
false,
false,
false,
false,
false
] | 0 | true | Requires two steps: 7×1.50=10.50, then 20-10.50=9.50. Base models often get the final subtraction wrong. | Qwen/Qwen3-1.7B-Base | 5 |
2 | logical_negation | All mammals are warm-blooded. Snakes are not mammals. Therefore, snakes are | not | [
"not warm-blooded.",
"not warm-blooded. Which of the following arguments has the same reasoning method?",
"not warm-blooded. What type of reasoning is this?",
"not warm-blooded. What is the fallacy in the above argument?",
"not warm-blooded. This argument is flawed because it commits which fallacy? Addition... | [
true,
true,
true,
true,
true
] | 1 | false | The logically valid conclusion is 'not necessarily warm-blooded'. Models frequently complete with 'cold-blooded' (which is true but a different claim) or incorrectly say 'warm-blooded' by following the pattern of the first sentence. | Qwen/Qwen3-1.7B-Base | 5 |
3 | temporal_date_reasoning | Today is Wednesday. The meeting was scheduled 5 days ago. The meeting took place on | friday | [
"the $m$-th day. What is the value of $m$?",
"what day of the week?",
"which day of the week?",
"what day?",
"what day of the week?"
] | [
false,
false,
false,
false,
false
] | 0 | true | Wednesday minus 5 days = Friday (previous week). Models often count forward instead of backward, or land on Monday by miscounting. | Qwen/Qwen3-1.7B-Base | 5 |
4 | spatial_reasoning | A red box is to the left of a blue box. The blue box is to the left of a green box. The green box is to the left of a yellow box. From right to left, the order is: | yellow | [
"red, blue, green, yellow. Which box is the closest to the yellow box? To determine which box is closest to the yellow box, we need to understand the given information and the order from",
"yellow, green, blue, red. What is the position of the red box from the left? The answer is:",
"red, blue, green, yellow. W... | [
true,
true,
true,
true,
true
] | 1 | false | Requires reversing a 4-item spatial chain. The correct order right-to-left starts with yellow. Models confuse direction or drop items. | Qwen/Qwen3-1.7B-Base | 5 |
5 | counterfactual_reasoning | In the real world, water freezes at 0°C. Imagine a fictional world where water freezes at 50°C. In this fictional world, a lake at 30°C would be | liquid | [
"considered too cold for swimming. How many degrees Celsius above the freezing point does this lake temperature represent?",
"equivalent to how many degrees Celsius in the real world?",
"considered \"freezing\" for most aquatic life. If a scientist measures the temperature of a lake in this fictional world and ... | [
false,
false,
false,
false,
false
] | 0 | true | 30°C < 50°C, so in the fictional world, the lake is liquid (not frozen). Models often apply real-world physics: 30°C > 0°C so 'liquid' — getting the right answer for the wrong reason, or saying 'frozen' by anchoring to 30°C sounding cold. | Qwen/Qwen3-1.7B-Base | 5 |
6 | low_resource_language | Swahili sentence: 'Mtoto anakula chakula.' English translation: 'The child is eating food.' Swahili sentence: 'Daktari anaandika barua.' English translation: | doctor | [
"'The doctor has given medicine.' Swahili sentence: 'Mwanamme cha nyingine.' English translation: 'I am tired.'",
"'The doctor is healing.' Swahili sentence: 'Mwanamme wameza kuwa kwa kazi.' English translation: '",
"'The doctor is reading a book.' Swahili sentence: 'Ndege nafasi cha mwanamke.' English translat... | [
true,
true,
true,
true,
true
] | 1 | false | 'Daktari anaandika barua' means 'The doctor is writing a letter.' Qwen3 is trained on 119 languages but Swahili coverage is sparse. The model may hallucinate a plausible-sounding but wrong translation. | Qwen/Qwen3-1.7B-Base | 5 |
7 | rare_entity_knowledge | The Tropic of Capricorn passes through the following continents: South America, Africa, and | australia | [
"Antarctica. Is this statement correct? Additionally, which of the following statements about the Tropic of Cap",
"which other continent?",
"Australia. The continent that the Tropic of Capricorn does not pass through is ____",
"South Asia. The Tropic of Cancer passes through the following continents: North Am... | [
false,
false,
true,
false,
true
] | 0.4 | false | The Tropic of Capricorn passes through South America, Africa, and Australia. Models sometimes add Asia or Europe, or stop at Africa. | Qwen/Qwen3-1.7B-Base | 5 |
8 | code_logic_tracing | x = 3
for i in range(4):
x = x * 2
# After the loop, x equals | 48 | [
"12.",
"24",
"24.",
"24",
"24"
] | [
false,
false,
false,
false,
false
] | 0 | true | 3 × 2^4 = 48. The model must mentally execute 4 iterations. Common errors: stopping after 3 iterations (x=24), or computing 3×4×2=24. | Qwen/Qwen3-1.7B-Base | 5 |
9 | consistency_under_rephrasing | The number of days in a leap year is | 366 | [
"not 365, but 36",
"____.",
"1/4 more than the number of days",
"________ the number of days in a common year",
"the same as the number of days in how many"
] | [
false,
false,
false,
false,
false
] | 0 | true | This should be trivial — 366. We run it 5 times to test consistency. Inconsistency (sometimes 365, sometimes 366) reveals poor calibration even on basic facts. | Qwen/Qwen3-1.7B-Base | 5 |
10 | compositional_analogy | Analogy: Painter is to Canvas as Sculptor is to | stone | [
"what?",
"___",
"?",
":",
"?"
] | [
false,
false,
false,
false,
false
] | 0 | true | The expected analogy: painter works on canvas, sculptor works on stone/marble/clay. The model may answer 'sculpture' (wrong category — product not medium), or 'chisel' (the tool not the medium), or 'gallery'. | Qwen/Qwen3-1.7B-Base | 5 |
README.md exists but content is empty.
- Downloads last month
- 4