load_dataset

### Describe the bug

I am performing two operations I see on a hugging face tutorial (Fine-tune a language model), and I am defining every aspect inside the mapped functions, also some imports of the library because it doesnt identify anything not defined outside that function where the dataset elements are being mapped:

https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb#scrollTo=iaAJy5Hu3l_B

`- lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
batch_size=batch_size,
num_proc=4,
)

- tokenized_datasets = datasets.map(tokenize_function, batched=True, num_proc=4, remove_columns=["text"])
def tokenize_function(examples):
model_checkpoint = 'gpt2'
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, use_fast=True)
return tokenizer(examples["text"])`

### Steps to reproduce the bug

Currently handle all the imports inside the function

### Expected behavior

The code must work es expected in the notebook, but currently this is not happening.

https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb#scrollTo=iaAJy5Hu3l_B

### Environment info

print(transformers.__version__)

4.46.1