Synthetic data designed as prompts for generating embeddings training data for retrieval. The "iteration" column refers to how the data was generated.

Iteration 1: Use the following pool of seed tasks, prompt GPT-3.5-Turbo to generate additional tasks.

RETRIEVAL_EXAMPLES = [
  'Provide a scientific claim as query, retrieve documents that help verify or refute the claim.',
  'Search for documents that answers a FAQ-style query on children\'s nutrition.',
  "Retrieve company's financial reports for a given stock ticker symbol.",
  "Given a book name as a query, retrieve reviews, ratings and summaries of that book.",
  "Search for scientific research papers supporting a medical diagnosis for a specified disease.",
  "Given a question, retrieve Wikipedia passages that answer the question.",
  "Provided a user question, retrieve the highest voted answers on Reddit ELI5 forum.",
  "Given a web search engine query, retrieve relevant passages that answer the query.",
  "Find Amazon reviews similar to the input review.",
  "Find the song lyrics most related to the user's search.",
  "Given a multi-hop question, retrieve documents that can help answer the question.",
  "Retrieve tweets that are semantically similar to the given tweet",
  "Given a news summary, retrieve other semantically similar summaries",
  "Given a question, retrieve relevant answers from Stackexchange",
  "Given a scientific paper title, retrieve paper abstracts that are cited by the given paper."
]

Iteration 2: Use the ~40,000 tasks generated in Iteration 1 as seed tasks, prompt GPT-3.5-Turbo to generate additional tasks.

Iteration 3: Use the ~80,000 tasks generated in Iterations 1-2 as seed tasks, prompt GPT-4-Turbo to generate additional tasks.