Datasets:

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

MMLU score recreation

#5
by theblackcat102 - opened

May I ask how do to recreate the MMLU score found in this figure? The score is much higher than what I can get using lm-evaluation-harness using the models from ablation collections. I wonder what do I get wrong

HuggingFaceFW org

I've added a note to the top of this file detailing how you can reproduce our setup: https://huggingface.co/datasets/HuggingFaceFW/fineweb/blob/main/lighteval_tasks.py

@guipenedo thanks for the quick reply, will close for now ~

theblackcat102 changed discussion status to closed

Sign up or log in to comment