Canonical train/dev/test splits
Hello
I had liked this dataset for future reference but then forgot about it. Now I stumble upon it again. This is great material for Dutch benchmarks! I think it'd be valuable to have stratified train/dev/test splits for the dataset to report in papers. Do you think you could add these to the dataset by splitting up the current training set? Probably preferably stratified on the answer column to avoid biases in the order of the answer. That would be an awesome addition! If you need help, I can assist where needed.
Thanks
Bram
Hi Bram,
Thanks for your interest! If benchmarking is your current goal, I would like to mention the Cohere INCLUDE benchmark (ICLR 2025), where this dataset is integrated.
https://huggingface.co./datasets/CohereForAI/include-lite-44
https://huggingface.co./datasets/CohereForAI/include-base-44
If you're still interested in a different pre-set train/dev/test split, I can have a look at it next week!
Best,
Mike
Aha, I did not know about that! Though the worry is, then, that people may find your dataset and train on it. But if a subset of yours is used as a test set in the benchmark, then that may lead to unfair gains in the result, because that test set is in this training set. So splitting up into train/validation/test at least (where validation and test items are the same ones as in the Cohere datasets) still seems useful I think. If you want I can also have a look and help out.
Agreed, I think the stratified way you mentioned is a good approach. If you have some time and could possibly create a pull request, that would be greatly appreciated! If not, I'll still have a look at it next week.