WTForbes/

nan_new

This release contains a cleaned full training set and a frozen evaluation set. Audio is embedded in Parquet and can be streamed directly by the Hugging Face datasets library.

train: 3,209,619 rows, 10475.161 hours eval: 709 rows, 1.976 hours

The train split contains 3,209,619 cleaned base rows. No external code-switch corpus is included in this release. The frozen eval audio is physically separated from… See the full description on the dataset page: https://huggingface.co/datasets/WTForbes/nan_new.