UNLP 2026
Ukrainian document question answering with retrieval and reranking
When Local Wins Fail to Generalize: UNLP 2026 Ukrainian Document QA
UNLP 2026 shared task, 2026
The task asks multiple-choice questions about a collection of Ukrainian documents, and systems have to run offline within a 9-hour compute budget. My pipeline retrieves pages with BGE-M3, reranks them with Qwen3-0.6B, and answers with a quantized MamayLM-Gemma-3-12B.
Results
- Private leaderboard score of 0.8722, 7th of 15 teams
- 0.933 top-1 document retrieval on the development set
- The tweaks that helped most on the development set did not survive the hidden test domains. The plain, conservative retrieval setup generalized better, which is where the title comes from.