HuggingFaceFW/fineweb · Details on the evaluation with lighteval

May 3

Hello! I have some questions about the evaluation of models.

for dataset siqa the metrics acc/acc_norm are not available in lighteval, just the ones from helm (exact_match, quasi_exact_match, prefix_exact_match prefix_quasi_exact_match), should I add the task or did you use the metrics from helm?
for mmlu did you average the accuracies of tasks from the leaderboard, the ones here, or do the results come from the single task mmlu of lighteval?
for arc are the results the average of arc-easy and arc-challenge?
are all the tasks 0-shot or few-shots?

Thank you for any help :)

HuggingFaceFW org May 4

Hi! I've just added our full list of lighteval tasks to the repo: https://huggingface.co/datasets/HuggingFaceFW/fineweb/blob/main/lighteval_tasks.py, you should be able to reproduce our results with this file.
Everything was 0-shot.

Let me know if you still have other questions :)

May 14

I could perfectly reproduce the results, now I will use the same evaluation for my experiments.
Thank you very much, very helpful :)

guipenedo changed discussion status to closed Jun 4