Correct minor errors in example notebooks for causal language modelling

[SumanthRH]

What does this PR do?

Corrects minor errors in dataset preprocessing in the example notebooks at examples/causal_language_modeling . I believe there are two mistakes that can cause issues when people are using the same code for a different model or task:

1. Input tweet text and labels are concatenated after tokenization. However, with 🤗 tokenizers, a BOS token may or may not be added depending on the model if no explicit value is passed for the add_special_tokens argument. For example, with GPT2, a BOS token is not added, while for Llama-2 a BOS token will be added. Because of this, if you simply tokenized the label with tokenizer(labels) and then concatenate with the input sequence, you can have a stray BOS token. You can quickly check this out yourself with the following code block:

from transformers import AutoTokenizer
gpt2_tokenizer = AutoTokenizer.from_pretrained("gpt2")
llama_tokenizer = AutoTokenizer.from_pretrained("stabilityai/StableBeluga2") # stable beluga has the same tokenizer as Llama 2
input_text = "@HMRCcustomers No this is my first job"
label = "Neutral"
input_formatted = f"Tweet text : {input_text} Label : "
concat_seq_llama = llama_tokenizer(input_formatted)["input_ids"] + llama_tokenizer(label)["input_ids"]
concat_seq_gpt2 = gpt2_tokenizer(input_formatted)["input_ids"] + gpt2_tokenizer(label)["input_ids"]
concat_seq_llama_decoded = llama_tokenizer.decode(concat_seq_llama)
concat_seq_gpt2_decoded = gpt2_tokenizer.decode(concat_seq_gpt2)
print(concat_seq_llama_decoded)
print(concat_seq_gpt2_decoded)

The outputs for Llama 2 and GPT are <s> Tweet text : @HMRCcustomers No this is my first job Label : <s> Neutral and Tweet text : @HMRCcustomers No this is my first job Label : Neutral respectively. In some cases, the BOS and EOS tokens are the same, so this can lead to lower performance.
2. The final token after concatenating the label should be an EOS token, which may be different from the padding token. Correct me if I'm wrong!

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[younesbelkada]

LGTM thanks, however I would like @pacman100 to have a look here if possible as he wrote those notebooks, just to be sure!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[pacman100]

Thank you @SumanthRH for fixing the example!