The Fifth Ukrainian NLP Conference (UNLP 2026) organizes a Shared Task on Multi-Domain Document Understanding. This Shared Task aims to challenge and assess AI capabilities to find the right information in a stack of domain-specific documents and generalize across domains.
Registration Form
Please register by February 21, 2026.
The Kaggle link will be sent to the registered participants on January 15, 2026.
Join the discussions in Discord via https://discord.gg/DYNnWaZD4a.
2025-12-20:
- Add more questions for the dev set: 40 => 461 questions.
- Postpone the Kaggle release till January, 2026.
2025-12-17:
- Add
readme-en.txtfiles describing the domains in English.
In this shared task, your goal is to build a solution that, given a set of documents, can (1) find the correct answer to a multiple-choice question and (2) identify the document and page where the answer was found.
The competition will run on Kaggle in the code-only mode, which means that you will submit your solution as code, e.g., a Python script or notebook, rather than submitting just test answers. Your solution should be able to run on a single P100 GPU and complete all test sets prediction within 9 hours. The code-only mode will facilitate fair competition. ⚖️
You are free to use any open-source embeddings and large language models of your choice — just keep in mind the compute limitations, which may not allow you to use the larger LLMs.
We encourage you to use LLMs specifically trained to work with the Ukrainian language, like LapaLLM and MamayLM.
Input data are grouped into domains. Each domain is represented by a folder of PDF documents with anonymized filenames and two readme files that describe the domain: readme.txt in Ukrainian and readme-en.txt in English.
The data is partitioned as follows:
15% — dev
25% — test_public
60% — test_private
Dev and test_public have two domains, while test_private has three domains. The additional domain will help assess how well your solution adapts to an unknown domain.
Please see the sample dev files for two domains in the data folder.
The score consists of two equal parts: the correct answer to a multiple-choice question and the correct reference (the document and the page in the document).
The final score is computed as:
where:
and the page proximity score is
Please see the sample questions for the dev files in the data folder:
- Your input data are columns
Question_ID,Question, andA-Fthat represent possible answers to the question. - Your solution should predict the values in
Correct_Answer,Doc_ID, andPage_Num. - We provide
DomainandN_Pagesof the correct document fordevto get you started. These columns will not be present in the test data.
Consider generating more questions on the dev set for self-assessment while we are working on the Kaggle setup.
Participants in the shared task are invited to submit a paper to the UNLP 2026 conference. Please see the UNLP website for details. Accepted papers will appear in the ACL anthology and will be presented at a session of UNLP 2026 specially dedicated to the Shared Task.
Submitting a paper is not mandatory for participating in the Shared Task.
- December 1, 2025 — Shared task announcement
- December 20, 2025 — Second call for participation
- January 14, 2026 — Third call for participation
- January 15, 2026 — (tentatively) The Kaggle competition link is sent to the registered participants
- February 21, 2026 — Registration deadline
- February 22 (11:58 PM GMT +02:00), 2026 — Final submission of system responses
- February 25, 2026 — Results of the Shared Task are announced
- March 8, 2026 — Shared Task paper due
- April 15, 2026 — Notification of acceptance
- April 26, 2026 — Camera-ready Shared Task papers due
- May 29–30, 2026 — Conference dates
Email: [email protected], [email protected], [email protected], [email protected]
