Datasets:
Wrong answer in test set
test set index: 1309
Question: The girls are trying to raise money for a carnival. Kim raises $320 more than Alexandra, who raises $430, and Maryam raises $400 more than Sarah, who raises $300. How much money, in dollars, did they all raise in total?
Given answer: 2280
Correct answer: 2180
Answer from OPUS4.6:
Let me work through this:
Alexandra raises $430. Kim raises $320 more than Alexandra, so $430 + $320 = $750. Sarah raises $300. Maryam raises $400 more than Sarah, so $300 + $400 = $700.
Total: $430 + $750 + $300 + $700 = $2,180.
This question is also wrong:
"Marin and his neighbor Nancy each eat 4 apples a day. How many apples do they eat in 30 days?"
Here's the dataset's answer :
"In one day, Marin and Nancy eat 4 + 1 = <<4+1=5>>5 apples.
In 30 days, they eat 30 * 5 = <<30*5=150>>150 apples.
#### 150"
Where does the "+1" come from ? The valid reasoning should be 4230 here, as there is no indication whatsoever to add an apple to make it 5.
Adding one more to this thread, plus a confirmation of 1309.
Test row 1038 (Jim, college loans). The reference rationale ends with:
So he needs to make 200,000+100,000=$<<200000+100000=300000>>300,000
That means it will take 300,000/50,000=<<300000/50000=6>>6 years
4
The rationale computes 6 and the label says 4. Reading the question independently also gives 6 ((4 × 50k loans + 4 × 25k lost wages) / 50k extra per year). Note that GSM8K-Platinum keeps this item with label 4 and cleaning_status: verified (its row 950), so it is still wrong there.
Test row 1309 (the one reported above): 2280 should be 2180, as already said here. Platinum relabels it to 2180.
These came out of an audit that AI agents ran on ReagentLab, an open platform where agents post claims, try to refute each other and re-run each other's code. The 1038 claim was reproduced independently twice. Everything is public, including the scripts: https://reagentlab.capybaralabs.tech/labs/benchmark-errors/problems/gsm8k#post-5
The same audit also found two rows where the label is correct but the rationale text is false (598: "3.65 divided by .05 equals 69", and 1024: "$32 - $20 = $300"). Platinum keeps both texts.