license: mit
Based on Sberquad
Answer converted to human affordable answer.
Context augmented with some pices of texts from wiki accordant to text on tematic and keywords.
This dataset cold be used for training retrieval LLM models or modificators for ability of LLM to retrieve target information from collection of tematic related texts.
Dataset has version with SOURCE data for generating answer with specifing source document for right answer. See file retrieval_dataset_src.jsonl
Dataset consists of 45278 examples in russian language of format:
{ 'text': 'text with correct answer', 'q': 'question text', 'a': 'correct answer text', 'context': 'text of 4-10 text chunks, one with right answer and others relevant with text and question on tematic and keywords' }
Length of one example of context + question + answer is less than 7000 symbols. It should be less than 2048 tokens of rugpt tokenizer.
File retrieval_dataset_src.jsonl has additionally SOURCE data for every text chunk in context, also SOURCE of right answer is set in answer. This variant of dataset is useful if you need extract answer with specifing source of the right answer.
{ 'text': 'text with correct answer', 'q': 'question text', 'a': 'correct answer text with SOURCE data of text', 'context': 'text of 4-10 text chunks, one with right answer and others relevant with text and question on tematic and keywords. Each of text chunks has it's own SOURCE data' }
All SOURCE data are sintetic generated and not real.