Datasets:
type mismatch in language_score
The language_score features of the DCLM and fwe3 subsets have mismatched data types (float32 vs. float64). This causes an error when trying to concatenate or interleave these subsets. It would be great if you could use a consistent data type for both (ideally float32).
Hello! Thanks for trying out Zyda-2!
For this release we decided to preserve original schemas/types of the component datasets. That's why in our instructions we recommend downloading our processed components separately, as they actually have different schemas. We recommend selecting only common columns, which are text and nemo_id, before concatenating datasets together.
Hi @yury-zyphra , I'm bumping into this problem too, when using shuffle() and streaming=True, it will access random parquet files and throw this error.
For future releases, if you could make a 10B and 100B token subset available for easy processing (like fineweb) but with the consistent data types it'd be very useful!
Hi
@alexjc
. Thanks for trying out our dataset! As I explained before, this is expected because we preserved original columns from the components, so that users can inspect the. For example, only DCLM and FineWeb-EDU have float quality scores, while for Dolma and Zyda quality score is categorical. You need to load each component separately, select only common columns nemo_id and text and then concatenate them together, if you want a single dataset. Also, if you simply concatenate, the weights would be suboptimal, I would recommend interleaving components using datasets.interleave_datasets() method (see our instructions).
I prepared a 100B tokens sample, that I've just uploaded. It only contains common columns and it was sampled with optimal weights
I updated the Download section in the Dataset card with examples.