Open deadmau5p opened 7 months ago
In this thread: https://www.kaggle.com/competitions/llm-detect-ai-generated-text/discussion/457889. The first point states that LLM/transformers do not work well for detection of ia generated text. I just do not know why.
https://www.kaggle.com/competitions/llm-detect-ai-generated-text/discussion/458466 another interesting discussion. Talking about encoder models not performing well. Possible reason would be that test data contains lots of noise, which trained model hardly undestands.
I should research for other encoder model experiments on kaggle to see if anyone got better results with Roberta-like models.