Check the user provided eos / bos token id against the tokenizer eos / bos token id

[ShashankMosaicML]

Currently, we need to specify the eos or bos token id in the dataset config for sequence id masking to work. This PR adds code to check if the provided eos and bos token ids match the token ids in the tokenizer. If there is a mismatch, an error will be raised. The error can be suppressed through the flags override_eos_token_id_mismatch_error and override_bos_token_id_mismatch_error

Verification runs:

125m-incorrect-eos-token-id-override-B1AOYT
125m-incorrect-eos-token-id-5DXPCI 
125m-correct-eos-token-id-Ge3qzT 

125m-incorrect-bos-token-id-override-mxBMzA
125m-incorrect-bos-token-id-EeBgKH
125m-correct-bos-token-id-I97jYG

[vchiley]

Should we also pull out 'bos_token_id'?

[ShashankMosaicML]

Should we also pull out 'bos_token_id'?

ConcatenatedSequenceCollatorWrapper throws an error if we set both the eos_token_id and bos_token_id. That is why I only set eos_token_id here. I can add flags that dictate which of these to automatically pick. But then, we will have to explicitly set one of those flags.

[vchiley]

@samhavens @codestar12 do you think eos_token_id or bos_token_id is more important and/or widely used?
(probably eos but wanted to check)

[samhavens]

@samhavens @codestar12 do you think eos_token_id or bos_token_id is more important and/or widely used? (probably eos but wanted to check)

Llama and t5 models both expect both eos and bos, OPT I think is BOS only?

[ShashankMosaicML]

@samhavens @codestar12 do you think eos_token_id or bos_token_id is more important and/or widely used? (probably eos but wanted to check)

Llama and t5 models both expect both eos and bos, OPT I think is BOS only?

OK, i think the safest thing to do would be to raise an error if the EOS or the BOS tokens that are provided in the yaml are different from what the tokenizer has. I will also add a flag that can override this error (in case someone wants to train with EOS/BOS tokens different from the ones in the tokenizer).

[vchiley]

Should we also warn/error if tokenizer does have eos_token_id or bos_token_id but user does not set it?