Aveni Labs has published the first instalment of our comprehensive safety and ethics framework for FinLLM, our suite of financial large language models. This governance report details the systematic approach we’ve developed to ensure responsible AI development and deployment in financial services.
What’s Inside
The report outlines our multi-layered governance structure, built around a dedicated AI Governance and Ethics Board that oversees every stage of the FinLLM development lifecycle. We detail how our framework integrates requirements from the EU AI Act, UK FCA and PRA guidelines, and cutting-edge ethical AI research from our University of Edinburgh partnership.
Key areas covered include:
Accountability Structure: Our tiered governance model spans from technical implementation teams to senior leadership, with clearly defined roles and responsibilities for ethical oversight, risk management, and regulatory compliance.
Base Model Selection: A systematic evaluation process that scores popular model families against 10 transparency criteria, including architecture disclosure, training data transparency, and responsible use guidance. The report includes our scoring results across major model families from Meta, Google, OpenAI, and others.
Documentation Framework: Comprehensive policies covering model cards, data protection impact assessments, risk management protocols, and privacy notices that ensure full traceability from core AI principles to technical implementation.
Industry Collaboration: Strategic partnerships with the FCA Digital Sandbox, regulatory compliance specialists, and research institutions that keep our governance approach current with evolving regulatory requirements.
This governance foundation enables us to develop FinLLM models that meet the complex regulatory demands of financial services while maintaining operational efficiency. The framework provides a blueprint for responsible AI development that balances innovation with robust safety measures.
The next instalment will examine our approach to data collection, training methodologies, and model evaluation processes.
