Create dataset loading script to fix loading
This HuggingFace Datasets upload is very broken due to differences in the data affecting normal loading.
After digging, I found out that "The type of metadata available varies from source to source" is a problem when using a HF dataset without a loading script. For example, some of the sources for 4chan_meta_sep have metadata fields that include truncated_portion, and some do not. This raises an exception when Datasets tries to load all the sources into a pyarrow table since the schemas of each source don't match.
Because of this, I don't think this dataset can be supported just with yaml metadata in the README.md, and instead it needs a dataset loading script. See these resources for more info:
- https://huggingface.co/docs/datasets/en/dataset_script
- https://github.com/huggingface/datasets/blob/main/templates/new_dataset_script.py
It's pretty frustrating that this dataset doesn't work since it seems really useful! I hope the AI2 team can test this public release and fix it soon so the rest of the research community can properly build on their work.
Sorry for your frustrations! Or original upload of this dataset is intended for use with the inference code linked in the readme (which consumes the jsonl.gz files directly). But I've updated the YAML now and changed how we represent the optional metadata fields so that the dataset now also can be loaded natively in HF datasets. Sorry for the slow response; for some reason I have not been receiving notifications for these community discussions.