How to create a "Both" option in multiple choice model training to solve IndexError: Target out of bounds error?

I was trying to train a pre-trained transformer model “bert-base-uncased”


So this is a model for multiple-choice questions where the prompt is combined with each option like prompt+optionA, prompt+optionB.

But if there is an option called “Both”, which has only a label say (2) but is not paired up like optionA and optionB, then the error will raise like IndexError: Target 2 is out of bounds.
Because I have 2 options but 3 labels. I can say that because If I eliminate the data containing labels 2 the training process goes fine.

The real case for me is:
There is a prompt and 2 responses from 2 AI models response_a and response_b for the same single prompt. (i.e. 2 options )
Now the model has to determine which response is better for this there are 3 labels winner_a, winner_b and winner_tie (i.e. 3 labels )

So how to deal with that? Should I create a third option? If so then how? If there is another solution please let me know.
Here is my tokenized function :

def tokenize_function(examples):
    # Repeat each prompt twice to go with the two possibilities of responses.
    first_sentences = [[context] * 2 for context in examples["prompt"]]

    # Pair the responses.
    second_sentences = [[a, b] for a, b in zip(examples['response_a'], examples['response_b'])]
    # print(second_sentences)

    # Flatten everything.
    first_sentences = sum(first_sentences, [])
    second_sentences = sum(second_sentences, [])

    # Tokenize.
    tokenized_examples = tokenizer(first_sentences, second_sentences, truncation=True)


    # Un-flatten.
    return {k: [v[i:i+2] for i in range(0, len(v), 2)] for k, v in tokenized_examples.items()}

here is the data structure :

    train: Dataset({
        features: ['id', 'model_a', 'model_b', 'prompt', 'response_a', 'response_b', 'winner_model_a', 'winner_model_b', 'winner_tie', 'class_name', 'labels', 'encode_fail', 'options', '__index_level_0__'],
        num_rows: 45981
    validation: Dataset({
        features: ['id', 'model_a', 'model_b', 'prompt', 'response_a', 'response_b', 'winner_model_a', 'winner_model_b', 'winner_tie', 'class_name', 'labels', 'encode_fail', 'options', '__index_level_0__'],
        num_rows: 11496