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

Area of Science:

• Biomedical Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Large language models (LLMs) show promise for biomedical named entity recognition (BioNER).
• Fine-tuning LLMs for BioNER demands significant computational resources, limiting accessibility.
• Parameter-efficient fine-tuning (PEFT) offers a solution for resource-constrained environments.

Purpose of the Study:

• Investigate PEFT techniques for optimizing LLMs in BioNER under limited computational conditions.
• Maintain competitive model performance and in-domain generalization capability.
• Explore multi-task BioNER capabilities of fine-tuned LLMs.

Main Methods:

• Utilized QLoRA, a PEFT method, to fine-tune the Llama3.1 model for BioNER, creating NERLlama3.1.
• Developed an LLM instruction tuning dataset using existing BioNER datasets (NCBI-disease, BC5CDR-chem, BC2GM-gene).
• Fine-tuned Llama3.1-8B on a single 16GB GPU and incorporated self-consistency NER prompting (SCNP) during inference.

Main Results:

• NERLlama3.1 achieved high F1-scores: 0.8977 (NCBI-disease), 0.9402 (BC5CDR-chemical), 0.8530 (BC2GM-gene).
• Demonstrated strong generalization on unseen datasets (e.g., 0.8378 on NLM-gene).
• Outperformed fully fine-tuned LLMs and BERT-base in BioNER tasks, showing superior in-domain generalization.

Conclusions:

• NERLlama3.1, using PEFT, achieved high BioNER performance with significantly reduced computational requirements.
• The model exhibits superior in-domain generalization compared to traditional pre-trained models.
• Low resource demands, high performance, and strong generalization make NERLlama3.1 suitable for diverse clinical BioNER applications.