Accounting-Instruction Tuning: Explainable and Regulation-Compliant Risk Detection in Financial Reports

Abstract

Large language models have transformed natural language processing across many domains, but their application to financial reporting - a field requiring strict regulatory compliance - remains limited. We propose Accounting-Instruction Tuning (AIT), a framework for aligning generative models with accounting standards including International Financial Reporting Standards (IFRS) and US-GAAP. Using 24k multilingual annual reports and 5k expert-curated excerpts for earnings management detection, we fine-tune LLaMA-2-13B and ChatGLM-6B with a dual-objective training approach. Our approach balances classification accuracy with explanatory transparency via a composite loss function. Evaluation on FinRisk and CN-Fraud datasets shows promising results: AIT achieves 10.8% higher F1 scores than FinBERT and outperforms zero-shot GPT-4 by 7.1%, while improving regulation-specific compliance by 12.9%. Generated explanations demonstrate strong faithfulness (FactScore fidelity: 0.82) and coverage (0.78), validated through both automatic metrics and expert assessment. Notably, our implementation requires only 48 hours of training on standard academic hardware (two A100-80GB GPUs), making this approach accessible to the broader research community.