Question about Validation Split and Coverage of Legal Texts

#10
by chmaycha - opened

Hello,

Thank you for providing this dataset! I am currently using it for continued pretraining of a language model and had a question regarding the validation split.

Is it possible that the validation set contains new legal texts or laws that are not present in the training set? If so, do you have any recommendations on how to best handle this scenario during pretraining, especially if we want the final model to benefit from all available legal content?

chmaycha changed discussion status to closed
chmaycha changed discussion status to open
CoAStaL NLP Group org

As mentioned in the dataset card, and the related article:

MultiEURLEX is chronologically split in training (55k, 1958-2010), development (5k, 2010-2012), test (5k, 2012-2016) subsets

For pre-training purposes, you should use merge all the splits in a single dataset, which can be easily done with the datasets.concatenate_datasets() function.

You can shuffle the dataset, or keep the chronological order, in case you want the model to be mostly informed by the most recent legislation.

kiddothe2b changed discussion status to closed

Sign up or log in to comment