(Georgia Institute of Technology, 2024-04-27)
Baheti, Ashutosh
As we increasingly rely on powerful language models, ensuring their safe and effective operation necessitates extensive research in controllable text generation. Existing state-of-the-art language models struggle to generate the most accurate or desired output at the first attempt. Inspired by recent developments in self-correction in large language models and new reinforcement learning methods, we aim to train smaller language models as fine-grained editors, whereby they iteratively edit outputs to satisfy threshold constraints over multiple classifier-based attributes.
In this thesis, I show a study of contextual offensive behavior of pretrained large language models and curate a high-quality dataset for toxicity detection. Next, I introduce a novel offline RL algorithm that can utilize arbitrary numeric scores as rewards during training to optimize any user-desired LM behavior by filtering out suboptimal data. Finally, I designed an offline RL framework, I propose a fine-grained multi-attribute controllability task, where the goal is to guide the language model to generate output sequences that satisfy user-defined threshold-based attribute constraints. The LM model can take multiple edits to reach the desired attributes. Experiments on both languages and proteins demonstrate the versatility and effectiveness of our approach.