Usage

C2D and D2C sysnthetic data are used to train the MiniCheck models from the work (GitHub Repo):

📃 MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents (link)

C2D: We start with any human-written claim statement. The goal is to generate synthetic documents that require models be able to check multiple facts in the claim against multiple sentences each.

D2C: We start with any human-written document to start with. The goal is to generate claims and pair them with portions of the human written document, which, once again, require multi-sentence, multi-fact reasoning to check the claims.