Instruction Finetuning Llama3-8B Model Using LoRA for Financial Named Entity Recognition

Battle Count: 74

Financial NER is a foundational component for quantitative trading pipelines. Accurate extraction of company names, monetary values, dates, and quantities from financial reports and news enables automated event-driven trading strategies, sentiment analysis, risk factor identification, and knowledge graph construction for market intelligence. However, the paper focuses on the NER task itself rather than direct trading applications, and the small dataset limits immediate production deployment.

LoRA fine-tuning is relatively straightforward with available libraries (PEFT, HuggingFace Transformers). The main complexity lies in: (1) data annotation and instruction formatting, (2) GPU memory management for 8B parameter models (though LoRA significantly reduces this), (3) hyperparameter tuning for the specific domain, and (4) evaluation pipeline setup. The paper uses modest hardware requirements (batch size 4, gradient accumulation 6, bf16 precision), making it accessible for smaller organizations.

The paper provides detailed hyperparameters (LoRA rank, alpha, learning rate, batch size, epochs, precision), dataset statistics, and instruction templates. However, no code repository or dataset link is provided. The dataset of 1,693 sentences is not publicly available. The instruction template is described but the exact training pipeline code is not shared.