subclassing_for_string_conversion.py

from transformers import AutoModelForSequenceClassification
import torch


class FineTunedDistilBertWithStringLabels(AutoModelForSequenceClassification):
    def __init__(self, config):
        super().__init__(config)
        self.label_dict = {
            0: "Clarification",
            1: "Factual",
            2: "Operational",
            3: "Summarization"
        }

    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, labels=None):
        # Perform the usual forward pass
        outputs = super().forward(input_ids, attention_mask, token_type_ids, labels)

        # Get logits (raw model output)
        logits = outputs.logits
        predicted_class_index = torch.argmax(logits, dim=-1).item()  # .item() -> extract regular python number

        # Map index -> string label using dictionary
        outputs.label = self.label_dict[predicted_class_index]
        return outputs