Multimodal LLM for Financial Analysis FinTral: Based on Mistral-7B Model, Scores Close to GPT-4
-
Recently, researchers from the University of British Columbia and Invertible AI introduced a groundbreaking large language model (LLM) — FinTral, specifically tailored for the financial sector. FinTral adopts a multimodal approach, capable of processing text, numbers, tables, and visual data to address the complexity of financial documents. The model introduces FinSet, a comprehensive benchmark for evaluating financial LLMs. Experiments demonstrate that FinTral excels in multiple tasks, including versions with enhanced visual and tool retrieval capabilities, surpassing established models like GPT-4 in numerous tasks.
FinTral is based on the Mistral-7b model and has undergone domain-specific pre-training using the FinSet dataset, which includes diverse sources such as C4, news, and financial documents, totaling 2 billion tokens, to improve its understanding and response capabilities for financial queries. To further enhance performance, FinTral has been optimized through instruction tuning and AI feedback combined with human and AI feedback. FinTral processes visual data via a CLIP encoder and handles numerical tasks through tools, effectively enhancing its functionality. The model further improves accuracy and depth in financial analysis through direct policy optimization and retrieval-augmented generation. Experimental results show that FinTral performs excellently in various financial tasks. The FinTral-INST model, through fine-tuning of pre-trained models, surpasses all other models with an average score of 0.49. Models enhanced with reinforcement learning from AI feedback also show significant progress, with FinTral-DPO achieving an outstanding average score of 0.59, just slightly below GPT-4's average score of 0.69. However, the results also highlight areas needing improvement, including real-time data processing, maintenance and updates, and the scarcity of annotated data.
FinTral is an advanced financial language model that leverages large datasets and diverse training methods to analyze complex financial data. By pre-training on clean financial data and employing retrieval methods, it reduces the risk of model hallucinations, improving accuracy and reliability. Its real-time adaptability to financial markets and dynamic data retrieval capabilities can significantly enhance prediction accuracy and decision-making. Researchers acknowledge the limitations and risk factors in the study but remain optimistic about future developments.
Paper URL: https://arxiv.org/abs/2402.10986