Datasets:

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
ArXiv:
DOI:
Libraries:
Datasets
Dask
License:

MMLU evaluation setting

#44
by jordane95 - opened

Hi, I noticed that you changed the format of MMLU from choice-based MCQs to the cloze QA format by log likelihood ranking, because the latter metric is more stable for small LMs. There could be a format where multiple choices are given in the context and the language model needs to select the most appropriate one by log likelihood ranking of the choice + full text. Have you tried using this metric?

HuggingFaceFW org

Hi, I know it's been many months but eventually we did try this. Surprisingly, it generally performed worse than either MCF or CF in our experience.

Sign up or log in to comment