Datasets:
Tasks:
Question Answering
Modalities:
Text
Formats:
csv
Languages:
Chinese
Size:
10K - 100K
License:
Did you use unfiltered TMMLU data?
#7
by
ZoneTwelve
- opened
If the datasets from TMMLU are used directly, it's possible that they might carry the same issues found in TMMLU.
I've realized from this discussion
Reference
We have filtered out any issues found in TMMLU which we have found in addition in our other sources:
- missing context dependencies : depends on previous question, missing question context (ie 題組 mentioned in issue 15), missing images
- invalid multi choice questions : questions with E option, non multi choice question, multi answer multi choice question ( contain more than 1 options, ie A,B )
I hope these covers what mentioned in the above issues. If you found any questions containing any of the above issues which we missed, we are happy to filtered it out and push a revised version.
Sounds great!!!
If I encounter similar issues with TMMLU+, I will let you guys know.
ZoneTwelve
changed discussion status to
closed
ZoneTwelve
changed discussion title from
Are you using unfiltered TMMLU data?
to Did you use unfiltered TMMLU data?