After the invention of ChatGPT, several language models offering dedicated services for different industries have been developed. BloombergGPT is one such generative AI large language model developed for the finance sector. This language model is trained with massive financial data to answer questions related to the finance industry.
Bloomberg, the leading financial service provider, invented Bloomberg GPT to offer AI-based services to its users. The company has trained this model with extensive data and machine-learning techniques to ensure it provides accurate information to financial advisors, experts, and investors. This post gives an overview of Bloomberg GPT and how it functions.
What is Bloomberg GPT?
Bloomberg GPT, an internal AI model, can more accurately answer questions like “CEO of Citigroup Inc?”, assess whether headlines are bearish or bullish for investors, and even write headlines based on short blurbs. The model is trained with a mixture of financial and generic datasets. The team has used a 50 billion parameter decoder-only language model to train Bloomberg GPT.
Bloomberg will have access to Bloomberg terminal, a software used by investors and financial experts to keep up with everything related to the finance sector. Some services offered by Bloomberg include streaming real-time finance news and offering research reports and financial advice. The company also claims that this new model will improve the existing NLPs and assist in various tasks.
Bloomberg Unveils a GPT Finance-Focused AI Model
Bloomberg announced the invention of its new model, Bloomberg GPT, on March 30, 2023. The company released a research paper with all details regarding its project. Bloomberg GPT fulfills the need for a dedicated AI tool to offer finance-related services.
Bloomberg GPT outperforms the existing models and can understand complex and unique terminology related to the finance sector. The model pulls data from Bloomberg’s existing financial database, including its 363 billion token dataset with financial documents written in English. This model is anticipated to improve the existing NLPs of Bloomberg and offer a better platform for its users.
Bloomberg GPT is capable of generating headlines for newsletters. During its testing, the company said that Bloomberg GPT used only a few examples to generate queries using Bloomberg Query Language (BQL).
What BloombergGPT Brings to the Finance Table
Bloomberg has been in the finance sector for several years. The company has always contributed to the sector by developing efficient and helpful tools powered by technology. Currently, Bloomberg is working hard on its new project Bloomberg GPT. However, with several AI tools already available in the market, the question is: how is Bloomberg a better model? What are the features that it offers?
Training and Parameters
Bloomberg GPT is trained for specific financial NLP tasks. It is not a generic model like ChatGPT. Bloomberg announced that the model uses a 363 billion token dataset of English financial documents and a 345 billion token public dataset. In total, Bloomberg is developed using over 700 billion tokens.
Machine learning engineers have also trained this model with 50 billion parameter decoder-only causal language. The massive dataset and training methods allow the model to outperform existing NLPs by large margins.
Bloomberg GPT is trained using an extensive dataset but lacks multilanguage functionalities. The training dataset of Bloomberg includes newspapers, journals, research papers, web-scraped financial documents, etc., in English. This limits the software’s ability to understand input in any other language besides English.
Biases and Toxicity
Companies have successfully developed AI tools like ChatGPT, Bloomberg GPT, etc. But the battle to train these tools to stop processing harmful content or hallucinating is ongoing. All companies, including Open AI and Bloomberg, struggle to train their models to provide relevant information without hallucinating.
Bloomberg users have to face the same issue while using the platform. The software may sometimes hallucinate or generate irrelevant responses. It may also process disallowed content.
Bloomberg GPT is a closed model. Bloomberg hasn’t revealed much information regarding this software except for its general information and training parameters. The company hasn’t declared the model size or the data used for this model.
According to sources, the model is anticipated to be trained with the sensitive information from Bloomberg for several years. Bloomberg aims to make Bloomberg GPT accessible to Bloomberg terminal users who subscribe to its services. In simple terms, the model won’t be accessible to its users as a standalone app but can be accessed via the Bloomberg terminal.
Comparison to other similar open models
Bloomberg GPT is equipped with the latest information and trends in the financial sector. It is capable of producing far better results compared to similar-sized models. For instance, compared to GPT-NeoX and FLAN-T5-XXL, Bloomberg GPT answers queries faster and more accurately.
Besides, the model gave a remarkable performance when tested for financial tasks like giving knowledge assessments, reading data, extracting financial information, and linguistic tasks.