[LLM] Add exception for using different language in eval dataset

[github-actions[bot]]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[peladodigital]

Could you add more details on how this would work?

[sgpyc]

This is to allow models intended for languages other than English. This change will enable potential submitter to use non-English versions of C4 as eval dataset in open submissions. There is no initial ckpt or eval target for non-English eval dataset and potential submitters need to come up their own, and specify in their submission, similar to what Google did with a 480B model to BERT in the open division.

[fujitsu-notsu]

The current rule about dataset (Section 6) states and allows submitters to choose any public dataset for open divisio as follows :

OPEN: Any public dataset may be used for training the model, however the evaluation data must be drawn from the benchmark dataset in a manner consistent with the reference.

However the word 'benchmark dataset' seems ambiguous. Does it mean a dataset that is defined at section 3 as reference or one that is chosed by the submitter for open division? If the former, rule fix may be needed.

[sgpyc]

My understanding of the quoted rule is the open division still uses the same eval dataset as the close division, hence this PR.