Below you will find a dataset consisting of news clippings collected from open sources, which describe an event and contain named entities. Each text is accompanied by a blank sentence where you need to add that named entity from the list of options. These options are extracted from the text and may be repeated. The correct choice extracted from any place in the text is considered correct, i.e. There may be several correct answers. The main thing is that the summary is meaningful.
You are provided with a dataset of 70,000+ texts, which are news clippings. All text examples were collected from open sources and then automatically filtered using QA systems to prevent obvious issues from entering the data set. The texts were then filtered by the IPM frequency of the words they contained and finally manually reviewed.
There are 3 files available in the archive:
- train and val – for training the model and setting parameters;
- test – a test file in which you need to make a prediction.
For verification, you need to upload a jsonl file to the platform, which contains the question id and the answer to it.
The quality of the solution is assessed using the F1 metric.
