Large-scale language models are now widely available thanks to the continued growth and development of artificial intelligence. Models such as ChatGPT, GPT4, LLaMA, Falcon, Vicuna, and ChatGLM have shown excellent performance on a variety of traditional tasks, opening up a world of opportunities for the legal community. However, gathering reliable, up-to-date, high-quality data is essential for creating large-scale language models. Therefore, creating an effective and efficient open source legal language model has become important.
Large-scale artificial intelligence model development is impacting several industries, including healthcare, education, and finance. These models have proven useful and effective in handling difficult problems and generating insightful data. In the legal field, on the other hand, the need for intrinsic relevance and accuracy demands thorough research and the creation of unique legal models. Law is crucial in shaping communities, regulating relationships and ensuring justice. Legal practitioners make prudent decisions, understand the law, and provide legal advice based on accurate and up-to-date information.
The nuances of legal terminology, complex interpretations and the dynamic nature of law create special problems that require professional solutions. Even state-of-the-art models like GPT4 frequently produce incredible results regarding hallucinations and legal issues. People often think that if they use their expertise in the relevant domain to improve their models, they will get better results. However, this is not the case as there are still many hallucinations and inaccurate results in early legal LLMs (LawGPT). At first, they understood the demand for legal LLMs in China. However, there were no commercially accessible Chinese models with more than 13 billion parameters at the time. Combining training data from sources such as MOSS and increasing the Chinese vocabulary improved the foundation of an economically viable model, OpenLLAMA. This allowed researchers at Peking University to build a base model for Chinese and add law-specific data to train his law model, ChatLaw.
The main contributions of this paper are:
1. A successful method to reduce hallucinations: They improved the model training procedure and included four modules during reasoning: “consultation”, “referencing”, “autosuggestion”, and “response”. We present a method to reduce Hallucinations occur less frequently because the vertical model and knowledge base are integrated through reference modules. Reference modules incorporate domain-specific knowledge into the model and use authoritative data from the knowledge base.
2. A model is trained to extract legal feature words from the user’s everyday language. Based on LLM. With the help of this model that recognizes terms with legal meaning, legal situations within user input can be quickly and effectively identified and analyzed.
3. A model that measures the similarity between the user’s normal language and a dataset of 930,000 relevant litigation texts is trained using BERT. This allows you to build a vector database to quickly search for works with similar legal contexts for additional research and citation.
4. Develop a Chinese legal exam to assess the dataset: Create a dataset to assess the legal expertise of Chinese speakers. They also created his ELO Arena scoring system to determine how well different models perform in legal multiple-choice tests.
He also pointed out that a single general purpose legal LLM may only work well in this area for some jobs. As a result, a number of models have been developed for different situations, such as multiple-choice questions, keyword extraction, and question answering. Using the HuggingGPT technique, I used his LLM at scale as a controller to manage the selection and deployment of these models. Based on each user’s request, this controller model dynamically selects the specific model to activate, ensuring that the best model for the task is used.
Please check paper and github link.don’t forget to join 25,000+ ML SubReddits, Discord channeland email newsletterShare the latest AI research news, cool AI projects, and more. If you have any questions regarding the article above or missed something, feel free to email me. Asif@marktechpost.com
🚀 Check out 100’s of AI Tools at the AI ​​Tools Club
Aneesh Tickoo is a consulting intern at MarktechPost. He is currently pursuing his Bachelor of Science in Data Science and Artificial Intelligence from the Indian Institute of Technology (IIT), Bhilai. He spends most of his time working on projects aimed at harnessing the power of machine learning. His research interest is in image processing and he is passionate about building solutions around it. He loves connecting with people and collaborating on interesting projects.
