Financial services are facing a new reality: generative AI is transforming the industry. Although closed models offer powerful functionality, they can raise concerns around transparency and control. This is where open source LLMs come into play. Now more advanced than ever, these models provide a flexible and secure foundation for creating AI-driven solutions. With open source, financial institutions can embrace innovation while maintaining control of their data and algorithms.

Gaurav Sharma, Client Partner, Financial Services at Fractal, spoke with AIM to explore this shift and explain how open source LLMs are used in financial services.

Sharma highlighted the critical importance of data protection when deploying open source LLMs. Companies must ensure strict compliance with data privacy laws and regulations. This involves minimizing data collection, filtering content appropriately, and deploying models locally wherever possible. Techniques such as anonymization, serialization, and differential privacy are essential tools for protecting sensitive information.

It developed a framework for managing data privacy: detect, process and rehydrate. Detection involves identifying potential risks to personal or sensitive information. Treatment addresses these risks through processes and governance structures, while rehydration focuses on integrating results into policy and governance.

Sharma emphasized the need for robust encryption protocols, data anonymization and comprehensive data governance policies to protect sensitive information.

The regulatory landscape presents significant challenges when using open source LLMs. Sharma identified four main concerns: data privacy, bias and fairness, explainability and scalability. “Data privacy involves processing sensitive information in accordance with regulations. Bias and fairness require taking into account ethical considerations and ensuring fair outcomes from the model,” Sharma noted.

Explainability is another critical area. “Regulations often require that model decisions be explainable in simple terms. The ability to articulate how models operate and make decisions is critical to compliance,” he added. Sharma also emphasized the need for scalability and efficiency, to ensure that LLMs can evolve with the needs of the organization while maintaining performance standards.

Bias mitigation and model explainability

Bias in open source LLMs is a serious concern. “To mitigate them, use diverse datasets and employ robust bias detection and mitigation techniques throughout the model lifecycle,” Sharma explained.

“Transparency and accountability are essential. Tools like LIME (Local Interpretable Model-agnostic Explanations) can help explain model decisions and build confidence,” he said.

Explainability is key to understanding and trusting open source LLMs. Sharma emphasized the use of tools and techniques that allow users to explore the model’s predictions and understand the results interactively. “Integrating attention mechanisms and saliency maps can provide insights into model predictions, making it easier to explain decisions in natural language,” he explained.

Performance and customization

On performance, Sharma acknowledged that while open source LLMs may initially lag behind proprietary models, they improve quickly. “Open source models like GPT-Neo, Mistral and Llama now perform at levels comparable to proprietary models. Companies are realizing the potential of these models, with usage shifting from 80-20 in favor of proprietary models to a more balanced 50-50,” he noted.

Customization is a significant advantage with open source LLMs. Sharma highlighted that open source models provide flexibility to tailor solutions to meet specific financial regulations and requirements. Customization helps refine models to meet unique financial needs and compliance standards. Although this requires resources, it provides the opportunity to develop highly specialized applications.

Hidden costs and security considerations

Sharma, however, warned of the hidden costs associated with open source LLMs. “The initial investment in IT resources, development and maintenance is significant. Personalization requires ongoing expertise and effort. Unlike proprietary models, which offer ready-to-use solutions, open source LLMs require significant in-house development,” he said.

Other costs include ensuring regulatory compliance and integrating models into existing systems. Sharma noted that open source solutions require careful consideration of security measures, compliance with regulations and integration into business processes. The need for continuous performance optimization and support adds to the overall investment.

Additionally, recruiting the right talent requires capital. After the initial investment in compute, companies need the right talent to maintain the system pipeline throughout adoption. “You have to have the right expertise to move forward. »

It is essential to prevent the misuse of open source LLMs. Sharma mentioned best practices to mitigate risks like malware and harmful content. Exposing LLMs to adversarial examples during training, implementing robust input validation, and controlling access are essential steps.

“Ensuring appropriate security environments and feedback loops helps protect against malicious activity,” he said.

Adoption and scalability

Sharma noted that when adopting open source LLMs, companies get scared, especially when the conversation around scalability accelerates. “As companies talk about generative AI and all the fancy things, they also want to talk about POC,” Sharma said.

Companies must be committed to advancing LLMs because every company has different needs and what suits them. “I think companies need to determine where they stand on several levels before moving forward,” Sharma added, adding that when it comes to the sustainability of these open source LLMs in the long term, they “learn and grow”.

Once a few companies begin to adopt a specific LLM, others begin to learn and grow along the way. “It requires a lot of investment, but then that investment will pay off.”

“Sustainability is crucial,” Sharma emphasized. “We are currently in the awareness phase, but assimilation and adoption will follow.”