Today marks an exciting milestone for Servient as we announce the release of InstructLAW, the first legal domain specific large language model. InstructLAW can be implemented within a law firm’s dedicated infrastructure making InstructLAW a core strategic asset of the firm.
Large language models (LLMs) and their implementations such as ChatGPT have ignited the imagination of many in the legal profession. ChatGPT has exposed the public to the power of LLMs.
LLMs have profound implications for the legal profession. Goldman Sachs recently predicted that 44% of all legal tasks will be automated through the use of LLMs. See Article
Servient has been working with language models since the introduction of BERT in 2018. Recent years have seen substantial and fast moving improvements in LLMs
Instruction-tuned LLMs are the current state of the art. An instruction-tuned LLM is fine-tuned on instructions and examples of successful attempts to follow the instructions. Instruction-tuned LLMs can be further fine-tuned through RLHF (Reinforcement Learning with Human Feedback) which improves the LLM by providing supervised feedback on the completed tasks.
What is InstructLAW?
InstructLAW is an instruction-tuned LLM that is trained on the language of the legal profession and fine-tuned to support legal tasks.
Every law student becomes immersed in the language of the law which then expands to the language of legal practice. The language of the legal profession is contained in legal cases, statutes, regulations, briefs, professional journals, research memos, contracts, patents and so on.
InstructLAW is trained on various legal sources. For example, InstructLAW has learned the structure of language from a dataset including over 10 million legal opinions and a huge variety of legal resources. Servient has implemented legal specific instruction tuning and RLHF to improve the model's performance on specialized legal tasks.
Why is a Legal Domain Specific LLM Important to the Legal Profession?
A Domain Specific LLM includes training on language specific to the domain and is fine-tuned for the domain tasks. When it comes to a legal task, a smart lawyer will perform the legal task better than an extremely smart person without legal training.
Client confidentiality is at the core of legal practice. InstructLAW addresses the dilemma that lawyers face with leveraging generative AI in the highly confidential setting of the attorney’s work. InstructLAW can be deployed within a law firm’s dedicated infrastructure allowing the lawyer to truly leverage the power of generative AI. As InstructLAW is under the control of the firm within their secure environment, the lawyer can use generative AI without fear of compromising client confidences.
Effective use of LLMs require sharing client information and attorney work product with the LLM. Even if the client information is not used to further train or fine-tune the LLM as promised by some API providers, confidential client information and attorney work product is nevertheless sent to the LLM through context prompting.
Instead, to derive the greatest benefit from generative AI, the lawyer should guide the LLM with their work product and client specific information. In fact, a number of legal tasks are best solved when the LLM is trained or fine-tuned on an attorney's work product.
InstructLAW will become a core asset of the firm as it is tuned by the work product of the attorneys. The legal knowledge and reasoning associated with the work product becomes an institutional asset of the firm. Much like investing in training new associates is a core need for the growth of a successful firm, the law firm of the future will invest in the institutional knowledge built into their dedicated InstructLAW model. Unlike lawyers who can move among firms, the law firm will not lose the institutional knowledge embedded in its InstructLAW model -- a significant core asset of the law firm of the future.
InstructLAW provides a cost efficient solution to operate at the scale required by legal tasks. When lawyers start to evaluate real production use of generative AI solutions they soon realize that metered API requests on a per token basis are very expensive when applied to true legal tasks. As InstructLAW is deployed within the dedicated infrastructure of the firm, there is no metered usage costs making production use of generative AI feasible for data-intensive workflows.
How is Servient Using InstructLAW Today?
Servient announced the release of Descriptive AI last summer. Descriptive AI allows the lawyer to enter a narrative description of the issues in the case or use existing work product such as a complaint or memo to initiate the machine learning workflow in eDiscovery. Descriptive AI relies upon InstructLAW to understand the lawyer's narrative and assist in the identification of documents relevant to the narrative.
Integration with InstructLAW allows lawyers to avoid the inefficient review of documents to train a TAR model. Generative AI allows the lawyer to begin the review process by interacting naturally with the software, explaining the nature of the claims and defenses in a narrative form; the AI takes it from there and identifies the documents relevant to the claims and defenses.
Descriptive AI is just one example of how Servient transforms legal workflows with generative AI. Stay tuned for more examples of legal workflows powered by InstructLAW. We can't wait to see what forward thinking lawyers will do with InstructLAW.