Datasets:
Multiple choice options
I saw on the web interface for humans to play the game that there are 4 options for each riddle. However, the dataset release only includes the gold reference answer. Would it be possible to publish the wrong options as well so that the benchmark can also be evaluated as a multiple-choice task by comparing loglikelihoods, rather than only by judging with another LLM?
Valid question @juliafalcao ! The wrong options are chosen from the wrong LLM answers that you can also find in the dataset. We considered a MCQ version but then did not pursue it because it adds the complexity of choosing distractor answers that are similarly difficult across riddles and communities. For the quiz it is for demonstration purposes but for loglikelihood LLM benchmarking it would need some refinement and review by humans with suitable guidelines.
Hi @JuliaKreutzerCohere Julia, Thank you so much for adding me to the Cohere Labs Community organization on Hugging Face. I appreciate the invite and I'm excited to explore and contribute. Please let me know if there's a recommended way to get started. Thanks again!
Valid question @juliafalcao ! The wrong options are chosen from the wrong LLM answers that you can also find in the dataset. We considered a MCQ version but then did not pursue it because it adds the complexity of choosing distractor answers that are similarly difficult across riddles and communities. For the quiz it is for demonstration purposes but for loglikelihood LLM benchmarking it would need some refinement and review by humans with suitable guidelines.
Thank you for the answer, @JuliaKreutzerCohere , I understand! We will discuss internally whether to select MCQ options from the wrong LLM answers or stick with an LLM-as-a-Judge approach.