Missing prompts

[Muennighoff]

Nice work! I think some of the prompts from your appendix are missing in DataLab?

Looking in this directory, I couldn't find e.g. prompts for sentence expansion like

source: TEXT: {summary} QUERY: What details would you include in a storyline to make it more
engaging and informative?

Is there a way to get all prompts you used similar to having all datasets in https://github.com/ExpressAI/DataLab/blob/main/datasets/rst/rst.py, so that one can easily reconstruct your full dataset for training? Thanks! 🤗

[Muennighoff]

Also is daily_mail_category supposed to work with the topic classification prompts? It sounds like topic clf to me, but doesn't seem to work 🧐

When running:

ds = load_dataset("rst", "daily_mail_category")
res = ds['train'].apply(template_tc2)
print(next(res))

I get

TypeError                                 Traceback (most recent call last)
[<ipython-input-39-3d757e349047>](https://localhost:8080/#) in <module>
      3 ds = load_dataset("rst", "daily_mail_category")
      4 res = ds['train'].apply(template_tc2)
----> 5 print(next(res))

[/usr/local/lib/python3.8/dist-packages/datalabs/arrow_dataset.py](https://localhost:8080/#) in apply_basic(self, func, prefix, num_proc)
    909         ]:
    910             for sample in self.__iter__():
--> 911                 labels = self._info.task_templates[0].labels
    912                 labels_to_answers = dict(zip(range(len(labels)), labels))
    913                 yield func(sample, labels_to_answers)

TypeError: 'NoneType' object is not subscriptable

cc @yyy-Apple @pfliu-nlp

[yyy-Apple]

Nice work! I think some of the prompts from your appendix are missing in DataLab?

Looking in this directory, I couldn't find e.g. prompts for sentence expansion like

source: TEXT: {summary} QUERY: What details would you include in a storyline to make it more
engaging and informative?

Is there a way to get all prompts you used similar to having all datasets in https://github.com/ExpressAI/DataLab/blob/main/datasets/rst/rst.py, so that one can easily reconstruct your full dataset for training? Thanks! 🤗

Yes, we have not added those prompts to DataLab yet, but the prompts we used for training are listed in full in the Appendix of the paper.

[yyy-Apple]

Also is daily_mail_category supposed to work with the topic classification prompts? It sounds like topic clf to me, but doesn't seem to work 🧐

When running:

ds = load_dataset("rst", "daily_mail_category")
res = ds['train'].apply(template_tc2)
print(next(res))

I get

TypeError                                 Traceback (most recent call last)
[<ipython-input-39-3d757e349047>](https://localhost:8080/#) in <module>
      3 ds = load_dataset("rst", "daily_mail_category")
      4 res = ds['train'].apply(template_tc2)
----> 5 print(next(res))

[/usr/local/lib/python3.8/dist-packages/datalabs/arrow_dataset.py](https://localhost:8080/#) in apply_basic(self, func, prefix, num_proc)
    909         ]:
    910             for sample in self.__iter__():
--> 911                 labels = self._info.task_templates[0].labels
    912                 labels_to_answers = dict(zip(range(len(labels)), labels))
    913                 yield func(sample, labels_to_answers)

TypeError: 'NoneType' object is not subscriptable

cc @yyy-Apple @pfliu-nlp

Currently, the RST datasets are not aligned with DataLab prompting functionality, we do not recommend using the apply(template) function.

[Muennighoff]

Currently, the RST datasets are not aligned with DataLab prompting functionality, we do not recommend using the apply(template) function.

So is it currently possible to recreate your dataset for training?