Dataload fix

[jahatef]

In #958

https://github.com/EleutherAI/gpt-neox/pull/979/files#diff-e2b248f8c422a601bcb0b7d93f96c1dff070f2694737e2b69f1def64ab9c1844L276

When data = None, _keys = ["text", "label"]

This led to access errors:

https://github.com/EleutherAI/gpt-neox/pull/979/files#diff-e2b248f8c422a601bcb0b7d93f96c1dff070f2694737e2b69f1def64ab9c1844R280-R287

where data_b does not have any values in the "label" key.

Instead,

https://github.com/EleutherAI/gpt-neox/pull/979/files#diff-e2b248f8c422a601bcb0b7d93f96c1dff070f2694737e2b69f1def64ab9c1844R306

was placed to only access the "label" key if neox_args.label_data_paths is set from the config file, which is the config determining if a labeled dataset is used.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
2 out of 3 committers have signed the CLA.

✅ StellaAthena
✅ Quentin-Anthony
❌ Quentin TastyRice

---

Quentin TastyRice seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[StellaAthena]

@Quentin-Anthony I have made the requested changes. I'm finding that training hangs in mulit-node settings with tp > 1, as reported in #985. However I see the same on main and don't think this PR is to blame.

[StellaAthena]

After applying this fix to both nodes it now works. Whoops xD