The ProofWriter dataset contains many small rulebases of facts and rules, expressed in English. Each rulebase also has a set of questions (English statements) which can either be proven true or false using proofs of various depths, or the answer is “Unknown” (in open-world setting, OWA) or assumed negative (in closed-world setting, CWA).

The dataset includes full proofs with intermediate conclusions, which models can try to reproduce.

The dataset supports various tasks:

• Given rulebase + question, what is answer + proof (w/intermediates)?
• Given rulebase, what are all the provable implications?
• Given rulebase + question without proof, what single fact can be added to make the question true?