Autogenerating a Domain-Specific Question-Answering Data Set to Enable High-Performing Language Models for Magnetic...

Domain-specific language models are attractive because they can be tailored to suit their users. However, the high-end computational resources needed to pretrain large language models (LLMs) by conventional methods tend to prohibit their development. Furthermore, there is a dearth of large, domain-specific, question-answering (QA) data sets to finetune LLMs for prompt engineering, even when computational resources can be found to pretrain LLMs. Our paper presents a method for autogenerating a data set of 168,080 domain-specific QA pairs relating to magnetic materials. We show how this can be employed to finetune LLMs for the downstream task of Question-Answering (QA) for the magnetic materials domain. We use the bidirectional encoder representations from transformers (BERT) architecture as our benchmark LLM to compare the relative performance of 6 LLMs. These include 3 vanilla (BERT-base-cased) LLMs that have been finetuned on various QA data sets (i) our domain-specific QA data set (MagQA); (ii) the Stanford Question-Answering Data set (SQuAD v2); (iii) our QA data set mixed with SQuAD v2. The other 3 (BERT-base-cased) LLMs have been subjected to domain-adaptive pretraining on a corpus of 97,308 magnetic papers prior to their finetuning on the same combination of QA data sets (i)-(iii). As well as our QA data set autogeneration pipeline, we show how our work can be extrapolated to the larger architectures of RoBERTa-base and DeBERTa-base. We also undertake an ablation study of the finetuning process using four sizes of QA data sets to gain insights into the quantity of data that are needed for sufficient domain-specification. The highest performing LLM was MagBERT_MagQA_Mixed, a vanilla BERT-base-cased LLM that had been finetuned on the combination of our autogenerated MagQA data set and SQuAD v2, achieving an F1 score of 78.43% and an exact-match score of 72.84% when tested on a manually annotated data set on magnetic materials. These results demonstrate that domain-specific BERT models only need to be finetuned from vanilla BERT models (i.e. no domain-adaptive pretraining is needed), pending the availability of sufficiently large, high-quality, domain-specific QA data sets that this work shows how to autogenerate.