wikipedia dataset loader needs to be updated to calculate stats for MEMIT and ROME

[KeremZaman]

https://github.com/zjunlp/EasyEdit/blob/2bbf0e1e878b355e77279e76fe1f167991a6f19e/easyeditor/models/rome/layer_stats.py#L102C1-L105C10

        raw_ds = load_dataset(
            ds_name,
            dict(wikitext="wikitext-103-raw-v1", wikipedia="20200501.en")[ds_name]
        )

20200501.en is no longer available in datasets library. So, it needs to be updated according to up-to-date usage (see https://huggingface.co/datasets/wikimedia/wikipedia):

ds = load_dataset("wikimedia/wikipedia", "20231101.en")

Since this is likely to affect results, it would be nice to have a way to use old wikipedia dataset too.

[JizhanFang]

Hello, we noticed that the file 20200501.en no longer exists at the path you provided. However, we found the same dataset at another location: ( https://huggingface.co/datasets/SamuelYang/wikipedia_20200501.en ). Additionally, we have some precomputed layer stats weight files corresponding to the following models: gpt-j-6B, llama2-7b, llama2-7b-chat, mistral-7b. I will upload the weights to the cloud drive and write a README file specifying the download links and the corresponding model files by the day after tomorrow.

[KeremZaman]

@JizhanFang thanks for your reply. Then, also it might be good to modify that line to allow users select SamuelYang/wikipedia_20200501.en as dataset in hparams file.