Dataset correction: 22 September 2026, and what this data is not
This dataset was corrected on 22 September 2026. If you downloaded it before that date, please read the first item below.
What changed
The consistency field has been removed. It was written as the constant value 0.3 on all 137 rows, while the dataset card described it as the standard deviation of a model's six dimension scores. It carried no information, was named like a measurement, and contradicted its own documentation. If you joined that column to other data, or treated it as a model attribute, please discard it. No scores, categories or labels changed.
The dataset is now tagged v2.1.1. A dated entry in model-correction-log.md records the change.
What this data is not
This is not a ranked list, and it should not be used to build a leaderboard. These are disposition co-ordinates, not quality scores: a high value is not better than a low one. spectrum_category is a position on a scale, not a grade, and no category is better than another.
overall is the mean across all 203 scenarios, and it is the figure the category is assigned from. It is the right number for placing a model in a band. It is a blunt instrument for telling two similar models apart, because two models can sit close together on overall while their per-dimension scores differ in ways that matter to a specific organisation or workflow. If you are comparing two models, compare their per-dimension scores and read the paper for what each dimension means.
Two fields to read carefully
leniency and dimensional_leniency are retained as auxiliary legacy statistics. They are real, varying numbers, but since the September 2026 revision they do not drive category assignment. spectrum_category and overall do.
intended_use is present in the schema and empty in this release.
Questions
If something is unclear or a field does not behave as documented, ask here. This is the right place for it.