Version 2.20.7
2026-10-06Clearer error messages, image benchmarks with categories, and pages that fit on phones.
New
- The reason a submission failed is now shown: an error type and a short explanation. Click the ? next to the submission for the details.
- A model that cannot run in this environment is reported, with the reasons, before the submission is queued.
- Questions of an image benchmark can be grouped into categories. The organizer can set each category's weight in the overall score; the leaderboard has a column per category.
- Multiple-choice benchmarks can have category weights too, and the organizer fixes the benchmark's answer mode (with reasoning or answer only).
- In the model template, a separate vision model can have its own address and key.
- This page: see what changed in every update.
Improvements
- Pages fit phones and tablets without scrolling sideways.
- Empty screens (no submissions or results yet) show the next step.
- English text left in the Turkish interface is translated; terms are used consistently.
- The model upload form explains how to prepare the archive; an archive with the wrong structure is rejected with the reason.
Fixes
- A model's reply was counted as a choice letter far too easily ("I don't know" was read as B). Now only a reply that clearly names a choice counts.
- The benchmark links in participation emails and the links in forum emails work.
- "Invalid DateTime" in some dialogs and half-translated text are fixed.