The Rise of Domain-Specific Models in Enterprise
Apr 24, 2024
Gradient Team
Not a One-Size Fits All
Most models that you hear about in the news are categorized as general-purpose models or foundational models. When it comes to these models, a massive dataset is often used to train these models - consisting of a wide array of source material that covers a broad range of knowledge. The output is a highly sophisticated general-purpose model that’s capable of executing a variety of language-based tasks, from simple text generation to answering questions and engaging in conversations.
However, while general-purpose models may be adequate for everyday tasks (e.g. responding to generic support questions or less complex tasks), a one-size-fits-all approach will likely fall short for enterprise businesses who are seeking industry-specific solutions that are capable of navigating industry specific needs.
Lack of Domain-Specific Knowledge: General-purpose models perform well across a broad set of applications, but they are often impractical for enterprise use cases that require domain-specific data and terminology. Without proper context or training, the model is likely to provide misleading or inaccurate responses - indicating a high probability of hallucination.
Inefficient Data Processing: In order for an LLM to provide accurate and contextual responses, it must be able to process domain-specific information efficiently and effectively. Given that general-purpose models are likely to lack context and relevant knowledge around the information that’s being processed, more often than not it will result in an increase in computational resources.
Advantages of Domain-Specific Models
Domain-specific models refer to models that are designed, trained, and optimized for a particular field or area of knowledge. This specialization enables the model to thrive in complex and highly regulated industries where precision is critical, supporting the teams and professionals using the technology (e.g. lawyers, medical providers, and investors).
In a recent study by Gartner, the research showed that by 2027 more than 50% of GenAI models that are used by enterprise will be domain-specific - focusing on either an industry or business function. This represents a 49 percentage point increase compared to 2023, reinforcing the need and desire for domain-specific models. So what do you get with domain-specific models?
Precision and Accuracy: By focusing on a narrower scope, these models can achieve higher accuracy and relevance in their responses, making them more reliable for domain-specific tasks. Domain-specific models are also trained on datasets that are rich in domain-specific information, terminology, and context - enabling the models to grasp the nuances of the specialized language and concepts in the field.
Efficiency and Optimization: Compared to general-purpose models, domain-specific models are more efficient when it comes to computational resources and response times. Since the model has been trained specifically for a purpose, it has the knowledge and context to efficiently and effectively handle domain related inputs.
Auditability and Citation: Given that most industries are heavily regulated, the need for auditability and proper citations are critical. Domain-specific models that are highly trained in their respective industries are able to breakdown and provide acceptable sourcing material when generating a response.
Are Domain-Specific Models Right For You?
Depending on your needs and the type of AI-driven solution you’re looking to build, you may not actually need a domain-specific model. The objective here is to select the type of model that will get you the best results, understand your field of work, and has the least amount of hallucination when generating a response. To determine whether a domain-specific model is right for you, you should consider:
Task Specificity: The more intricate and complex the task, the more likely you’ll need a specialized model. Ask yourself whether your task requires nuanced understanding or particular knowledge around a specific field.
Industry Complexity: Some industries can be complex and difficult to navigate, due to a variety of variables. Ask yourself whether or not your industry has nuanced processes, industry-specific terminology, industry-specific regulations, industry-specific compliance, ethical considerations, etc. If the answer is yes, you’ll most likely want to invest in a specialized model that has the right context and knowledge to provide appropriate responses.
Performance Comparison: Side by Side
Lets take a look at a real world scenario in which a domain-specific model would be best suited for the task vs. a general-purpose model.
When it comes to financial technology, NLP can be used to make use of unstructured data. A finance-specific LLM such as Gradient’s Albatross model, has been trained on the necessary skills and knowledge to execute against finance-related tasks such as: Classifying Financial Documents, Entity Recognition, Tabular Understanding, etc.
Deep Dive: Tabular Understanding
To see why a domain-specific model is needed, let’s dive into an example of tabular understanding. Understanding and being able to manipulate tabular data is critical for applying LLMs in an industry that runs on numbers. Much of the relevant information from unstructured documents in the industry is contained in tables. To be able to extract this information, LLMs must comprehend their tabular structure and perform numerical reasoning. Additionally, to accomplish any realistic task, downstream manipulation and inference is required after retrieval.
As a simple example, consider a investment decision that depends on projected, discounted cash flows pulled from a company’s 10K annual report. This requires finding and filtering a relevant table, forecasting a trend from this data, appropriately discounting and aggregating forecasted values, and finally suggesting an investment decision. A financial analyst would use a combination of prior knowledge in how SEC tables are structured as well as formulas in spreadsheets to arrive at an answer. Both intrinsic knowledge to interpret and manipulate a cash flow table as well as the ability to perform calculations are required. Neither are sufficient by themselves.
In our internal testing, we find that leading, general purpose models are unable to perform tasks of equal or lower difficulty to the example above. The table below shows accuracy on table summarization tasks - asking a generalist model to interpret and compute percentage change and sums - from tables generated from SEC documents. GPT-4 achieves only 50% accuracy in computing a percentage change across a dimension. As for V-Alpha-Tross, an earlier version of Gradient's Albatross model with limited capabilities available on Hugging Face, landed a perfect score on table extraction.
Getting Started with Domain-Specific Models
There are essentially 3 ways that enterprise businesses can get started with domain-specific models, each with it’s own set of challenges:
Option 1- Build From Scratch: While you can most certainly build your own domain-specific model from scratch and control the type of data that it’s trained on, it will require a vast amount of work just to train the model up to par with general-purpose models. If you choose to go with this option, you’ll need to keep in mind that it will require an enormous investment in time, budget, and resources.
Option 2 - Start with a Foundational Model: To avoid starting from scratch, the obvious option would be to build on top of an existing foundational model of your choice (e.g. Llama 3) and optimize the model further using domain-specific knowledge (e.g. fine-tuning, RAG, etc.). However this process can take anywhere from 6-15 months and can require up to $4m in development depending on the requirements.
Option 3 - Leverage an Out-of-the-Box Domain-Specific Model: The easiest option would be to leverage an out-of-the-box domain-specific model that can immediately meet your requirements. You can even supercharge the model to make it uniquely yours, by leveraging platforms like Gradient that enables you to further train these domain-specific models using your private data. This allows you to create a custom model that specializes in your industry and understands the ins and outs of your organization - providing the competitive advantage that you may be looking for.
How Gradient Can Help
The Gradient AI Foundry is the most comprehensive solution for AI transformation, providing compute infrastructure, domain and task-specific models and necessary tooling required for companies that are working on automation systems or operational tooling. By not having to worry about setup or infrastructure, enterprise organizations are able to use that time for other priorities.
Domain-Specific LLMs
Financial Services: Gradient provides finance organizations with a SOC 2 Type 2 AI copilot for their research and operational workflows. Albatross, Gradient’s proprietary state-of-the-art finance LLM (Large Language Model) and embeddings model, provides a strong base of financial understanding in order to create sophisticated solutions for finance related tasks.
View some of our customer stories on how Albatross is being used today to power financial use cases:
AML: Using AI to identify potential risk, detecting 3-4x more instances and reducing false positives by +50%
Customer Service Co-Pilot: Using AI to automate triaging and provide frontline support, resulting in 5-10x increase in productivity.
Investment Performance: Using AI to enhance market analysis and company evaluation to increase investment predictions by +20%.
Healthcare: Gradient provides healthcare organizations with a HIPAA-compliant AI copilot for their existing healthcare workflows. Nightingale, Gradient’s proprietary state-of-the-art healthcare LLM (Large Language Model) and embeddings model, provides a strong base of medical understanding in order to support and streamline AI-powered solutions within healthcare.
Take a look at how healthcare organizations are using Nightingale today and the success that they’ve seen:
Clinical Transcriptions: Using AI to automate clinical transcription, increasing transcription accuracy by 25%.
Healthcare Data Management: Using AI to harness full potential of their data to increase data coverage by 95% and accuracy by 87%.
Medical Benefits: Using AI to develop a medical benefits chatbot to provide 24/7 support and reduce operational costs by 70%.
Task-Adapted Multimodal LLMs
Aside from domain-specific models, Gradient offers task adopted multimodal LLMs. These models are built from the ground up and fine-tuned, in order to maximize performance across each use case. Today Gradient offers a variety of task-adapted multimodal LLMs including:
Data Summarization
PDF Extraction
Audio Transcription
Sentiment Analysis
Entity Extraction
And Many More
All Gradient Task-Adapted Multimodal LLMs are fully managed and accessible via an easy-to-use API, making it the easiest solution to integrate AI into your business workflows.
Want to get started with Gradient? Reach out to us!