Who can participate?
Anyone - students, researchers and practitioners, from any institution or country. The only exclusion is people who have had access to the full dataset.
Is there a registration fee?
No fee is charged to participate in the shared task. Any details about conference attendance for presenters will be announced.
Can I be on two teams?
No. Each person may belong to only one team. Duplicate registrations can lead to disqualification of all teams involved.
How is the 8B limit counted?
Add up the parameters of every neural component your system loads at inference, generators, ensemble members, classifiers, routers. MoE models count total parameters, not active ones. LoRA adapters count on top of their base model. Quantised models count their original parameters.
Does a LoRA adapter count towards the limit?
Yes, count the base model plus the adapter parameters.
Can I use a model released on 1 October 2026?
No. Only open-weight models released before 1 October 2026 are allowed.
Can I use GPT, Claude or Gemini?
Only to generate training data, and you must disclose it in your manifest and report. They may never be used at inference.
Can my system use retrieval?
No. Retrieval-augmented generation, knowledge bases, document stores, search indexes and lookup tables are not permitted at inference, whether online or stored locally. Your system must answer from the model's weights and the question itself.
Can my public leaderboard score different from my final score?
Your Codabench score comes from the full hidden test set. After the test phase closes, we reproduce each team's selected run from the submitted code; runs we can reproduce go into the final ranking.
No matching questions. Try another word, or ask on the forum.