Aligning Large Language Models (LLMs) with human feedback is crucial for their development. Existing preference optimization methods such as DPO and KTO, while improved based on Reinforcement Learning from Human Feedback (RLHF), are inherently derived from PPO, requiring a reference model that adds GPU memory resources and relies heavily on abundant preference data. Meanwhile, current preference optimization research mainly targets single-question scenarios with two replies, neglecting optimization with multiple replies, which leads to a waste of data in the application. This study introduces the MPPO algorithm, which leverages the average likelihood of model responses to fit the reward function and maximizes the utilization of preference data. Through a comparison of Point-wise, Pair-wise, and List-wise implementations, we found that the Pair-wise approach achieves the best performance, significantly enhancing the quality of model responses. Experimental results demonstrate MPPO’s outstanding performance across various benchmarks. On MT-Bench, MPPO outperforms DPO, ORPO, and SimPO. Notably, on Arena-Hard, MPPO surpasses DPO and ORPO by substantial margins. These achievements underscore the remarkable advantages of MPPO in preference optimization tasks.
Due to the scarcity of review data and the high annotation cost, in this paper, we primarily delve into the fine-tuning of pretrained models using limited data. To enhance the robustness of the model, we employ adversarial training techniques. By introducing subtle perturbations, we compel the model to better cope with adversarial attacks, thereby increasing the stability of the model in input data. We utilize pooling techniques to aid the model in extracting critical information, reducing computational complexity, and improving the model’s generalization capability. Experimental results demonstrate the effectiveness of our proposed approach on a review paper dataset with limited data volume.
This paper presents a study on the language understanding of the Dravidian languages. Three specific tasks related to text classification are focused on in this study, including abusive comment detection, sentiment analysis and fake news detection. The paper provides a detailed description of the tasks, including dataset information and task definitions, as well as the model architectures and training details used to tackle them. Finally, the competition results are presented, demonstrating the effectiveness of the proposed approach for handling these challenging NLP tasks in the context of the Dravidian languages.