README.md

metadata

dataset_info:
  features:
    - name: task_type
      dtype: string
    - name: instruction
      dtype: string
    - name: target
      dtype: string
    - name: source
      dtype: string
  splits:
    - name: train
      num_bytes: 89199350
      num_examples: 53264
    - name: validation
      num_bytes: 11738463
      num_examples: 6850
  download_size: 44500608
  dataset_size: 100937813
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*

Description

This is a dataset created for training Russian-language Seq2Seq and CLM models primarily for tasks related to Closed-Domain QA.

The dataset includes 3 main tasks:

1. AAQG (Answer-Aware Question Answering) - generation of questions based on context, provided the answer is known
2. QG - generating questions based on context, without a known answer
3. QA - the standard task of answering a question based on context.

AAQG, QG, QA tasks are generated based on regular datasets for which the context, question and correct answer are known. They are generated in a ratio of 0.4, 0.3 and 0.3, respectively.

List of datasets used to compile this dataset:

1. sberquad
2. russian_super_glue/muserc
3. russian_super_glue/danetqa

Prompts used for QA tasks:

AAQG_PROMPT = """AAQG | Текст: '{context}'. Ответ: '{answer}'"""
QG_PROMPT = """QG | Текст: '{context}'"""
QA_PROMPT = """QA | Текст: '{context}'. Вопрос: '{question}'"""

Authors

• Sergey Bratchikov/@hivaze (Tochka Bank)