Version 2.20.5
2026-07-21LLM judge evaluation, invitation codes, credits, benchmarks from Excel, and a broad security update.
New
- LLM judge (LLM-as-Judge) evaluation: in multiple-choice and question-answer benchmarks the model's answers can also be scored by a judge model.
- Invitation codes: administrators create and manage them, users can request one.
- Credits: every submission uses credit, and users can request more.
- Create multiple-choice and question-answer benchmarks from an Excel file, with a preview.
- Image benchmark type.
- Login with SSYZ / e-Devlet.
- Submissions run one after the other; waiting ones show as "Queued".
- Leaderboard sorting, filtering and download as CSV or JSON.
Improvements
- Large model files are uploaded in parts, more reliably.
- The whole interface is translated to Turkish; error pages are in Turkish.
- Login, session and file upload security strengthened; an account is locked for a while after many failed logins.
Fixes
- A stuck submission no longer blocks the queue; it is recovered automatically.
- Files with Turkish characters in their names download correctly.