Resource-efficient instruction tuning of large language models for biomedical named entity recognition

Objective:

Large language models (LLMs) have exhibited remarkable efficacy in natural language processing (NLP) tasks, with fine-tuning for Biomedical Named Entity Recognition (BioNER) receiving significant research attention. However, the substantial computational demands associated with fine-tuning large-scale models constrain their development and deployment. Consequently, this study investigates parameter-efficient fine-tuning (PEFT) techniques to optimize LLMs for BioNER under limited computational resources. By leveraging these methods, competitive model performance is maintained while preserving in-domain generalization capability.

Methods:

In this study, we employed the PEFT method QLoRA to fine-tune the open-source Llama3.1 model, developing the NERLlama3.1 model specifically designed for the BioNER task. First, an LLM instruction tuning dataset was created using BioNER datasets such as NCBI-disease, BC5CDR-chem, and BC2GM-gene. Next, the Llama3.1-8B model was fine-tuned using the QLoRA method on a single 16GB memory GPU. Furthermore, during the inference phase, we introduced a prompt engineering technique called self-consistency NER prompting (SCNP). This approach leverages the diversity of outputs generated by LLMs to significantly enhance NER performance. Finally, we also developed a multi-task BioNER-capable model, NERLlama3.1-MT, to investigate the capability of fine-tuned LLMs in addressing multi-task BioNER scenarios.

Results:

The NERLlama3.1 model achieved F1-scores of 0.8977, 0.9402, and 0.8530 on the NCBI-disease, BC5CDR-chemical, and BG2GM-gene datasets, respectively. Furthermore, when evaluated on previously unseen datasets, it attained F1-scores of 0.6867 on BC5CDR-disease, 0.6800 on NLM-chemical, and 0.8378 on NLM-gene. These results demonstrate that NERLlama3.1 not only outperforms fully fine-tuned LLMs but also exhibits superior in-domain generalization capabilities when compared to the BERT-base model. Additionally, this work represents the first exploration of fine-tuning LLMs for multi-task BioNER.

Conclusion:

NERLlama3.1 outperformed LLMs fine-tuned with full parameter updates, despite requiring significantly fewer computational resources. Moreover, it exhibited substantially superior in-domain generalization capabilities compared to traditional pre-trained language models. Its low resource demands, high performance, and strong generalization enhance its applicability and utility across diverse clinical BioNER tasks.