This repository contains datasets and scripts related to the paper:

Knowledge-Instruct: Effective Continual Pre-training from Limited Data using Instructions.

• data/ - Contains the datasets:

  • fictional_companies_questions_open_end.jsonl
  • popqa_v2.jsonl

  The context columns within these files contain the actual underlying data.

• create_fictional_companies.py - A script that outlines the process of generating the companies dataset underlying data.

• generate_open_end_questions.py - A script that outlines the process of generating the companies dataset questions and answers.

• Companies Dataset: Contains fictional company data, including question-answer pairs for reference.
• PopQA: This dataset is derived from the original PopQA but retains only tail knowledge.

To generate the companies dataset, first modify LLMOpenAI to your own version, then run:

python create_fictional_companies.py

Similarly, after changing LLMOpenAI, create question answer pairs by running:

python generate_open_end_questions.py