Percentage point formatting
In question 95, where you instruct to "Give your answer as a percentage to the nearest whole number", the answer is 50%
But in question 89 and 90, you said "what was the percentage difference" and "Give an answer as a percentage to 1 decimal place" consecutively. the answer is in range of 0 to 1 (0.318 and 0.304), making whenever LLM gives the answer 31.8% and 30.4% consecutively, becomes incorrect instead.
(Same problem goes for question 86, 87)
Can you standardize the answer format to put the same as question 95?
I don't think it makes the LLM incorrect, because the judge model handles that discrepancy
I don't think it makes the LLM incorrect, because the judge model handles that discrepancy
But the judge model can’t handle this correctly.
Take this query as an example:
"For the technology that accounted for the highest percentage of renewable energy generation, how much larger in percentage points was its contribution to total renewable generation in 2023 compared to the technology with the lowest contribution? Give an answer as a percentage to 1 decimal place."
The model’s response was:
"Based on the Clean Energy Australia 2024 report (Document 1), the technology that accounted for the highest percentage of renewable energy generation in 2023 was wind power, contributing 33.9% (Page 10, Page 85). The technology with the lowest percentage of renewable energy generation was bioenergy, contributing 3.5% (Page 10, Page 58).\n\nThe difference in percentage points is:\n$33.9\% - 3.5\% = 30.4$ percentage points\n\n30.4%"
The corresponding label is 0.304, yet the score is 0, which is clearly incorrect.
This suggests that a more capable judge model should be used, rather than Qwen3-235B-A22B-2507 (Non-Reasoning), or alternatively, the label itself needs to be corrected.
So far, there has been no official explanation for this issue.