Fine-tuned Language Models are Continual Learners

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6107 - 6122 December 7-11, 2022 ©2022 Association for Computational Linguistics Fine-tuned Language Models are Continual Learners Thomas Scialom1∗ Tuhin Chakrabarty2∗ Smaranda Muresan 2 1Meta AI 2Department of Computer Science, Columbia University tscialom@fb.com, tuhin.chakr@cs.columbia.edu, smara@cs.columbia.edu Abstract Recent work on large language models relies on the intuition that most natural language pro- cessing tasks can be described via natural lan- guage instructions and that models trained on these instructions show strong zero-shot perfor- mance on several standard datasets. However, these models even though impressive still per- form poorly on a wide range of tasks outside of their respective training and evaluation sets. To address this limitation, we argue that a model should be able to keep extending its knowledge and abilities, without forgetting previous skills. In spite of the limited success of Continual Learning we show that Fine-tuned Language Models can be continual learners. We empir- ically investigate the reason for this success and conclude that Continual Learning emerges from self-supervision pre-training. Our result- ing model Continual-T0 (CT0) is able to learn 8 new diverse language generation tasks, while still maintaining good performance on previous tasks, spanning in total 70 datasets. Finally, we show that CT0 is able to combine instructions in ways it was never trained for, demonstrating some level of instruction compositionality.1 1 Introduction Recent work has shown that large language mod- els have the ability to perform zero-shot and few- shot learning reasonably well (Brown et al., 2020; Rae et al., 2021; Smith et al., 2022). A particu- larly successful line of work relies on the intuition that most natural language processing tasks can be described via natural language instructions. For example, a summarization task can be reformat- ted as a response to a natural language input as shown in Table 1. Sanh et al. (2022) and Wei et al. (2022) have released T0 and FLAN respectively and shown that fine-tuning models on a massive ∗Both Authors Contributed Equally 1Our code is publicly available at https://github.com/ ThomasScialom/T0_continual_learning. The picture appeared on the wall of a Poundland store on Whymark Avenue [...] How would you rephrase that in a few words? Graffiti artist Banksy is believed to be behind [....] Table 1: An instance from T0 training set (Sanh et al., 2022) where a summarization task is reformatted as a natural language response to a natural language input. mixture of NLP datasets expressed via such natural language instructions (i.e., instruction tuning), im- proves the zero-shot performance of large language models. FLAN is extremely large in size (137B) and is not publicly available limiting its further use and reproducibility. Conversely T0 (Sanh et al., 2022) is publicly available and orders of magnitude smaller and hence we resort to working with T0. However impressive, these models are still lim- ited to simple instructions and mainly Natural Lan- guage Understanding (NLU) tasks. These models perform poorly on a wide range of tasks that differ from their respective evaluation sets. …

Content truncated to 500 words