Search

Reinforcement learning from human feedback (RLHF) has been extensively employed to align large language models with user intent. However, proximal policy optimization (PPO) based RLHF is occasionally unstable requiring significant hyperparameter finetuning, and computationally expensive to maximize the estimated reward during alignment. Recently, direct preference optimization (DPO) is proposed to address those challenges. However, DPO relies on contrastive responses generated from human annotator and alternative LLM, instead of the policy model, limiting the effectiveness of the RLHF. In this paper, we addresses both challenges by systematically combining rejection sampling (RS) and DPO. Our proposed method, RS-DPO, initiates with the development of a supervised fine-tuned policy model (SFT). A varied set of k responses per prompt are sampled directly from the SFT model. RS-DPO identifies pairs of contrastive samples based on their reward distribution. Finally, we apply DPO with the contrastive samples to align the model to human preference. Our experiments indicate that our proposed method effectively fine-tunes LLMs with limited resource environments, leading to improved alignment with user intent. Furthermore, it outperforms existing methods, including RS, PPO, and DPO., Comment: 16 pages, 4 figures

Large-scale conversational assistants like Alexa, Siri, Cortana and Google Assistant process every utterance using multiple models for domain, intent and named entity recognition. Given the decoupled nature of model development and large traffic volumes, it is extremely difficult to identify utterances processed erroneously by such systems. We address this challenge to detect domain classification errors using offline Transformer models. We combine utterance encodings from a RoBERTa model with the Nbest hypothesis produced by the production system. We then fine-tune end-to-end in a multitask setting using a small dataset of humanannotated utterances with domain classification errors. We tested our approach for detecting misclassifications from one domain that accounts for <0.5% of the traffic in a large-scale conversational AI system. Our approach achieves an F1 score of 30% outperforming a bi- LSTM baseline by 16.9% and a standalone RoBERTa model by 4.8%. We improve this further by 2.2% to 32.2% by ensembling multiple models., Comment: Accepted to ACL Findings 2021