Dataset Used for Reinforcement Learning Tasks of Tibetan Large Models

Over recent years, large language models exemplified by ChatGPT, operating on a prompt-based paradigm, have seen considerable advancements. These models are primarily built through predictive vocabularies, utilizing fundamental loss functions such as cross-entropy, without directly integrating human preferences or individual perspectives. To address this, a method known as Reinforcement Learning from Human Feedback (RLHF) has been devised. This approach fine-tunes language models with human over-sight, aiming not only to enhance their ability to execute a variety of written instructions more rigorously but also to generate content that is more accurate, logical, and reader-friendly. Currently, research on large Tibetan language models is still in its nascent stage, hindered by a scarcity of extensive datasets for pre-trained and fine-tuned models. This study introduces a dataset specifically designed for the reinforcement learning tasks of a large Tibetan model. It comprises two segments: 2000 manually crafted data points and 4000 automatically generated entries. Each data piece includes input for the model, the optimal output it should learn, and examples of suboptimal output. Utilizing this dataset, the model can identify areas for improvement and strive to produce higher-quality text.