[OpenVINO] Add gsm8k as a dataset option for CausalLM quantization

[nikita-savelyevv]

What does this PR do?

In this PR "gsm8k" dataset option is added to apply data-aware quantization to causal language modes. Calibrating on "gsm8k" in some cases provides better results compared to "wikitext2":

Model	Precision	Calibration Dataset	seqlen	gsm8k_cot_llama (strict-match)	gsm8k_cot_llama (flexible-extract)
Llama-3.1-8B-Instruct	nf4 per-channel	wikitext2	32	78.70%	78.77%
Llama-3.1-8B-Instruct	nf4 per-channel	gsm8k	64	80.29%	80.82%
Llama-3.1-8B-Instruct	nf4 per-channel	gsm8k	128	79.30%	79.53%
Llama-3.1-8B-Instruct	nf4 per-channel	gsm8k	256	81.20%	81.50%

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[nikita-savelyevv]

@ljaljushkin, please take a look.

[ljaljushkin]

Thanks, Nikita! LGTM

[IlyasMoutawwakil]

LGTM ! isn't the perf increase to be expected since you are calibrating on on test data/distribution ?

[nikita-savelyevv]

LGTM ! isn't the perf increase to be expected since you are calibrating on on test data/distribution ?

That's a good question! To be honest some experiments show higher accuracy when calibrating on wikitext2, so that seems not to be the case here. For example:

Model	Precision	Calibration Dataset	Seq. length	gsm8k_cot_llama (strict-match)	gsm8k_cot_llama (flexible-extract)
Llama-3.1-8B-Instruct	int4	wikitext2	32	83.17%	83.32%
Llama-3.1-8B-Instruct	int4	gsm8k	256	81.50%	81.80%

[rkazants]

Is it possible and reasonable to specify several datasets (e.g., wikitext2 and gsm8k) simultaneously for calibration? Can it provide higher accuracy?

I think it is reasonable to update documentation and mention in what cases to use gsm8k dataset.

[nikita-savelyevv]

Is it possible and reasonable to specify several datasets (e.g., wikitext2 and gsm8k) simultaneously for calibration? Can it provide higher accuracy?

Thanks for the idea! In theory it could, however there were no experiments done in this direction.

I think it is reasonable to update documentation and mention in what cases to use gsm8k dataset.

Experiments are still work-in-progress in order to determine when exactly it is beneficial to use gsm8k dataset. For now, the results are mixed, however in some cases gsm8k provides better quality. With this PR we add another option for users to try improve quantized model quality.