permit opt tokenizer

[bmosaicml]

What does this PR do?

Add functionality to support ICL with OPT tokenizer (which prepends special tokens each time tokenizer is called). Due to the way we concatenate context and continuation encodings into a single input, it's important that no special tokens are put between them.

GPT Neo 1.3 B gets following accuracy
metrics/piqa/5-shot/InContextLearningMultipleChoiceAccuracy: 0.717934787273407
metrics/lambada/0-shot/InContextLearningLMAccuracy: 0.5883721113204956

OPT 1.3B gets:
metrics/piqa/5-shot/InContextLearningMultipleChoiceAccuracy: 0.719565212726593
metrics/lambada/0-shot/InContextLearningLMAccuracy: 0.5883721113204956

MosaicGPT 1.3B gets:
metrics/piqa/5-shot/InContextLearningMultipleChoiceAccuracy: 0.637499988079071
metrics/lambada/0-shot/InContextLearningLMAccuracy: 0.4075581431388855

What issue(s) does this change relate to?

ICL eval doesn't work with OPT tokenizer since it adds special tokens between contexts/continuations.

Before submitting

• [ x] Have you read the contributor guidelines?
• [ x] Is this change a documentation change or typo fix? If so, skip the rest of this checklist.
• Was this change discussed/approved in a GitHub issue first? It is much more likely to be merged if so.
• [x ] Did you update any related docs and document your change?
• [x ] Did you update any related tests and add any new tests related to your change? (see testing)
• [ x] Did you run the tests locally to make sure they pass?
• [ x] Did you run pre-commit on your change? (see the pre-commit section of prerequisites)

[dakinggg]

LGTM, see the one comment

[abhi-mosaic]

@bmosaicml aren't the OPT scores a bit low? From our internal data, I see:

OPT 1.3B piqa 0-shot (not 5-shot): 0.7236
OPT 1.3B lambada 0-shot: 0.588

Vs:

OPT 1.3B gets:
metrics/piqa/5-shot/InContextLearningMultipleChoiceAccuracy: 0.626630425453186
metrics/lambada/0-shot/InContextLearningLMAccuracy: 0.37538760900497437

[bmosaicml]

@bmosaicml aren't the OPT scores a bit low? From our internal data, I see:

OPT 1.3B piqa 0-shot (not 5-shot): 0.7236 OPT 1.3B lambada 0-shot: 0.588

Vs:

OPT 1.3B gets:
metrics/piqa/5-shot/InContextLearningMultipleChoiceAccuracy: 0.626630425453186
metrics/lambada/0-shot/InContextLearningLMAccuracy: 0.37538760900497437

yes they are a bit low....where are you getting your numbers from?

Blocking until the low numbers are resolved

[bmosaicml]

@dakinggg @abhi-mosaic Sorry, the initial numbers I wrote were for OPT 125m I just retested with OPT 1.3B and the numbers match what we expect!

[dakinggg]

Fantastic! re-approving :)