pszemraj/riddlesense_plusplus

RiddleSense ++: Evaluating LLMs' Riddling abilities

Some cleaning and other modifications to make the riddle_sense dataset more suitable for training language models to generate riddles via standard text generation

Notable changes

• reformatting to use special tags indicating the question and answer components
• normalization: whitespace normalization, etc with clean-text
• spell correction the original dataset has numerous spelling errors; these are fixed using BertChecker from the neuspell library for the questions only. The answers were left as-is as there is some specific terminology/puns that get changed by SC.

Notes & Gotchas

• the original dataset has a non-commercial clause, so this does too.