This Polish-language dataset contains human-annotated question-answer pairs based on the 'Did You Know' (Czy wiesz?) format. It focuses on a binary classification task where models predict answer correctness against hard negatives chosen for their high token overlap with the source question.
Use Cases
- Train a binary classification model to predict answer validity based on the relationship between the question and the provided answer text.
- Benchmark Polish NLP models on their ability to resist lexical overlap bias using the hard negative samples.
- Develop information retrieval ranking algorithms that prioritize semantic relevance over simple keyword matching.
Strengths
- Human-annotated question-answer pairs specifically for the Polish language.
- Includes a q_id column to track and reference unique question instances.
- Negative answer candidates are selected based on the highest token overlap with the question to create challenging distractors.