Finetuning RoBERTa on SentiHood Dataset producing Random Outputs

Code: [ Google Colab ]

Is there any mistake in my usage of AutoModelForSequenceClassification or of the HF Trainer?

The data is conditioned for the Binary Classification task of predicting Yes/No for three sentiment classes (None, Neg, Pos) with different aspects and target_entities.

This approach was inspired by this paper: [ ]