Datasets:

wenge-research

/

yayi2_pretrain_data

like 50

Thanks for the great work on sharing the pretraining data of Yayi-2! It is of diverse sources and great text quality.

However, we found it quite difficult to categorize the data, because the title key in the JSON is somewhat arbitrary. There are source names such as arxiv, github, and refined-web, and there are real titles, e.g., Thomas Bradwardine.

In the techincal report, it stated

Through normalization, all raw data are formatted as JSON with keys such as data source, identifier, and content. Additionally, a language detector model is employed for language detection.

I wonder if is possible to share these pieces of meta-information, which can really facilitate our research on data ablation.

Thanks in advance!

Best,