InstrDialog, InstrDialog++
收藏资源简介:
本研究基于SuperNI数据集,创建了两个持续学习任务流:InstrDialog和InstrDialog++,分别包含19和19+19个任务,涵盖对话生成、意图识别等多个领域。数据集通过精心挑选的任务,旨在系统研究持续学习方法的效果。InstrDialog++进一步扩展了任务类型和领域,以探索不同类型任务对持续学习的影响。这些任务均通过自然语言指令模板转化为统一的文本到文本格式,便于模型理解和处理。数据集的应用领域广泛,旨在解决持续学习中的知识遗忘和知识转移问题,推动该领域的研究进展。
This study develops two continual learning task streams, InstrDialog and InstrDialog++, based on the SuperNI dataset, comprising 19 and 38 tasks respectively, covering multiple domains such as dialogue generation and intent recognition. Built with carefully selected tasks, this dataset aims to systematically investigate the effectiveness of continual learning methods. InstrDialog++ further expands the task types and domains to explore the influence of different task types on continual learning. All tasks are converted into a unified text-to-text format via natural language instruction templates, which facilitates model understanding and processing. The dataset has wide application scenarios, targeting at solving the issues of catastrophic forgetting and knowledge transfer in continual learning to advance research progress in this field.

- 1CITB: A Benchmark for Continual Instruction Tuning悉尼科技大学 · 2023年



