I want to investigate parameter-efficient fine-tuning (PEFT) methods (LoRA, bottleneck adapters, etc.) in the context of generative LLMs in different domains. I started reading the PEFT literature to find established benchmarks for my project. I saw people using datasets like SQuAD, E2E dataset, and XSum. Despite addressing multiple domains, there are no tags for the domain of each sample. I would need to have this information for my project. I could just use one dataset as one domain but the datasets I found do not usually have specific domains but contain samples from different domains. To summarize I would need datasets that

  • require a generative model (e.g. question answering with open answers, not multiple-choice)
  • cover a specific domain (sports, medicine, science, etc.) or contain this information as a feature for every sample


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.