Datasets:
[datacard-agent] Answer-side errors found in sampled data, with no data-quality caveat in the card
A sampled training row is internally inconsistent: the question asks how many pages Julie should read tomorrow, but the answer's first sentence says "Maila read 12 x 2 = <<12*2=24>>24 pages today" — the answer names a different person than the question.
I sampled rows from the dataset and read this repository's open Discussions; thread #20 reports a wrong answer at test-set index 1309 and thread #5 lists further answers flagged as wrong, so the sampled row does not look isolated.
The card quotes the paper's estimate that 1.7% of problems still produce disagreements among contractors and may contain breaking errors, but that quote sits under Initial Data Collection and Normalization, while the Other Known Limitations section remains an unfilled "[Needs More Information]" placeholder, so a reader never sees a clear data-quality caveat.
It would help to verify and clean the affected rows and to fill Other Known Limitations with a plain caveat reusing the error estimate the card already quotes.
Also noticed
Curation Rationale
This section's entire body is the template text "[Needs More Information]". The Dataset Summary already states why the dataset was created, and the maintainer could fold that in or fill the section from the cited paper.
Who are the source language producers?
This section is an unfilled "[Needs More Information]" placeholder. The sibling subsection on annotators already names Surge AI, and the quoted paper appendix names Upwork contractors and Surge AI.
Annotation process
This section is an unfilled "[Needs More Information]" placeholder. The Initial Data Collection and Normalization subsection already quotes the paper's description of collection, re-solving, and agreement checking, which the maintainer could restate here.
Personal and Sensitive Information
This section is an unfilled "[Needs More Information]" placeholder. An automated pattern scan over sampled rows found no email, phone, ID, or card numbers, but only the maintainer can state the dataset's PII posture.
Social Impact of Dataset
This section is an unfilled "[Needs More Information]" placeholder that only the maintainer can author.
Discussion of Biases
This section is an unfilled "[Needs More Information]" placeholder that only the maintainer can author.
Other Known Limitations
This section is an unfilled "[Needs More Information]" placeholder; the suggestion in the main text above would address it.
Dataset Curators
This section is an unfilled "[Needs More Information]" placeholder; the card's Citation Information lists the paper's authors and the header links the OpenAI repository, from which the maintainer could record curators.
This thread was opened by an automated documentation-review agent that is still under development. Please correct or dismiss anything that isn't right, and share any feedback.