The Pros and Cons of Using Domain-Specific LLMs for Financial AIs

You might have heard that there’s a choice that needs to be made when picking a large language model (LLM) for your AI. In truth there are dozens of choices, but the one that’s being talked about a lot in late 2023 is whether or not to use a domain-specific LLM.

The answer to that question depends on what vertical you’re coming from, and the amount of data available in your specific field. But we can tell you this: When building a financial AI, there are a lot of benefits to using a finance LLM.

This article will go through some of those benefits, while also addressing the most common drawbacks of taking such a specific route. By the end of it you should understand the pros and cons of using domain-specific LLMs for financial AIs, and be able to choose the most reasonable course of action for your business.

How Are Domain-Specific LLMs Different from Other Language Models?

‍

Domain-specific LLMs are trained with highly curated data sets, pulled from sources related to the specific industry that they will be serving. Unlike general large language models which take a more generic approach to the data ingestion process, domain-specific LLMs are designed to be used in AI that is being built for the targeted industries.

The impact on chatbots is a higher accuracy rating when conducting conversations in its field of expertise. For example, a general chatbot might assume that ‘AML’ stands for acute myeloid leukemia, which is both the number one search result and the most common medical result for that acronym. A financial AI trained with a domain-specific finance LLM would assume ‘AML’ refers to anti money laundering practices.

The impact on generative content includes the superior contextual understanding stated above, and more. Associated images, citations, and even the tone of voice used in an article or video will more closely match industry norms.

What is a Finance LLM?

Anyone who is looking for a large language model that will be relatable to customers in the financial field needs to avoid datasets that were trained ‘broadly’. When an AI ingests badly curated data, it makes bad assumptions on topics ranging from abbreviations and acronyms, to the experience level of both novices and experts in the chosen field, to the conversational tone they should be using. That’s where finance LLMs come in.

A finance LLM will include hundreds of millions, if not billions of parameters that define the resources that it should favor, the communications it should undertake, and the ‘on the job’ learning that it should conduct. Every parameter is a mini-rule that makes AI interaction easier and more accurate. But of course, the more complex the model, the higher the computational cost.

The most successful first-generation finance LLM was BloombergGPT. As you might have guessed from the name, this was a generative pre-trained transformer (GPT) that was designed for generative artificial intelligence. It was a great start, but arguably a little too broad. With over 700 billion tokens in its training set and a 50 billion parameter decoder, the team may have gone too broad and the end result might have been a bit too bloated. Still, it was the best first-generation finance LLM by a large margin.

An example of a next-generation LLM for financial AIs is the Alli Finance LLM. In order to offer the customer more flexibility as far as budget, computing power, and complexity are concerned, Allganize offers two different models of their finance LLM, with 13 billion and 70 billion parameters respectively.

Flexibility in the curation and training process of the Alli Finance LLM means that clients can choose between efficiency and complexity. The more streamlined model will offer lower computing costs, while the more robust model will offer higher levels of accuracy and relevancy.

The Pros of Using Domain-Specific LLMs for Financial AIs

• Chatbots that are more in tune with finance customers: This goes beyond using the right acronyms in a conversation. Financial AIs need to understand that they are not there to offer financial advice. Any AI that oversteps and implies that it is undertaking a fiduciary duty to the client is a disaster waiting to happen. A domain-specific LLM will allow for the training of financial AIs that understand their own limitations, and act as a guide to information rather than an interpreter of financial data.
• Generative AIs that resonate with their target audience: Financial AIs that need to generate content need to reflect the vocabulary and tone of their readers. This is called establishing the AI’s ‘voice’. These voices are far more appealing and understandable when a domain-specific LLM is used. Financial AIs are also expected to transcend ‘folk wisdom’ and common misconceptions about the financial world, instead offering solid, statistically backed data.
• Answers that are based more in fact and less in supposition: The ‘hallucination’ factor is well known to those who work with language models. It happens when an AI takes a bunch of ‘facts’ (some of which are not checked very well) and draws a wildly wrong conclusion. Financial AIs are far less prone to this, since their data ingestion range is from a far more narrow scope than a generalist AI.

The Cons of Using Domain-Specific LLMs for Financial AIs

• Danger of narrower points of view: Because everything the AI knows is heavily steeped in the world of finance, it might not understand business and lifestyle choices that are rooted in anything other than hard numbers. Whereas a general AI will be exposed to a lot of different corporate structures and voices, finance tends to be more of a monoculture outside of certain variables (risk appetite, investment strategy, and the like).
• Less reuse value in other sectors of the business: Clearly a large corporation that does more than finance would need to invest in either a general LLM or multiple domain-specific LLMs if they want to cover more than the financial world with their AIs.  

Summing It All Up

The main benefit of using a finance LLM is the specificity of the training that the AI will receive. It will understand financial shorthand, context, and user requests far more clearly than an AI using a more general LLM.

If you have any questions about which finance LLM to use, contact Allganize and consult with our team of experts. They’ve been able to advise several clients as to the best way to harness the power of AI for their businesses.