05 October 2023
The rapid advancements of Generative AI in the financial sector have ignited discussions about the suitability of different AI solutions for various firms. While Large Language Models (LLMs) like GPT-3 have received considerable attention for their impressive capabilities in text generation and language understanding, the often-overlooked Small Language Models (SLMs) offer unique advantages worth exploring. In this article, we will delve into the benefits of SLMs and their relevance in the financial industry.
Large language models such as GPT-3 undoubtedly exhibit remarkable capabilities across a broad spectrum of applications, including chatbots, search engines, and even coding tasks. However, their utilisation comes with several challenges. LLMs demand extensive computational resources and vast datasets for training, making them costly to run and challenging to deploy in production environments. These limitations raise questions about their suitability for some types of tasks.
Financial institutions and software providers are increasingly considering SLMs for tasks like contract analytics due to several key advantages:
SLMs are more lightweight and be more easily trained/fine tuned for specific use cases with one’s own data. It is much quicker and cheaper to fine tune a SLM than LLM (hours versus months) and in the evolving world of finance, where contracts have hundreds of amendments and regulations are introduced continually, this flexibility is a valuable asset.
Small language models are also much less compute intensive compared to their larger counterparts. LLMs require considerable GPU compute resources in order to run the models making them uneconomical for many scenarios. Whereas an SLM will run on very modest GPU hardware. For financial professionals, especially those dealing with real-time data, the efficiency of small models can be a game-changer.
Smaller models tend to exhibit reduced bias compared to their larger counterparts since they have fewer parameters and are less likely to capture biases present in training data. This feature is valuable in applications where fairness and bias mitigation are critical.
Smaller models may pose fewer privacy concerns compared to large models. They are less likely to memorise sensitive data from the training set, making them potentially more suitable for privacy-conscious applications.
One of the biggest challenges with LLMs is the so-called “long tail” problem. This refers to the fact that LLMs are typically trained on a huge amount of general-purpose data, which means they can be good at generative tasks such as answering general questions and generating text. But for more specialised tasks, commonly found in contract analytics, they can struggle to provide accuracy required. This is because the training data does not have enough examples of specific terms and language patterns that are common in contracts. As a result, LLMs can end up providing less accuracy than a SLM when trained on the same data set.
Smaller models are easier to fine-tune and customise for specific tasks or problem domains. This enables developers to create models tailored to their unique requirements. Where a firm holds a large number of bespoke clauses and terms for example, a SLM can prove to be an easily moldable solution, tailored to a specific problem/task.
So what does this mean for you? Where your competitors may be rushing to invest in large language models, perhaps consider the benefits of alternative AI or machine learning solutions that suit your business needs.
In an industry where precision, compliance, and efficiency are paramount, SLMs strike the balance between advanced automation and human judgement. They empower financial institutions to navigate the complexities of financial contracts with agility, confidence, and cost-effectiveness.
