Rocket AI

Pre-trained large language models (LLMs) like GPT-3 and LLaMA have demonstrated general capabilities for solving a variety of tasks. However, recent studies have shown that the abilities of LLMs can be further adapted for specific goals. Our approach leverages instruction tuning to adapt general-purpose LLMs for financial sentiment analysis, enhancing their understanding of numerical values and context for this particular task. The process involves transforming the sentiment analysis from a classification task to a text generation task, better aligning with the capabilities of LLMs. We then use the transformed dataset to instruction fine-tune the LLMs in a supervised manner. Finally, during inference, we map the generated outputs to sentiment labels.

Our specialized model, Rocket-AI, for financial sentiment analysis has undergone professional and accurate fine-tuning. Through analyzing and training our focused mini-models, derived from over 100,000 cryptocurrency-related events and event-driven quantitative back testing models, we provide expert interpretations of every incident occurring in the cryptocurrency market. This helps reduce barriers to understanding for users and assists them in making quick judgments and trades.

Figure 4: Usage of Rocket AI

Rocket-AI is the information analysis core of Rocket-3 with two key capabilities. First, it filters and analyzes Signals collected from Alpha Collection, generating Alpha insights. After Alpha generation, Rocket-AI conducts further analysis, curating truly valuable core Signals and Events.

The second function targets user Intention - Rocket-AI comprehends user demands, calculates prospective transaction paths for desired token purchases, and creates initial Proposals. Subsequently, it Estimates the Proposals by validating paths and analyzing feasibility.

Based on back-testing, Rocket-AI currently chooses optimal paths with over 80% accuracy. Before reaching 100%, all data is manually reviewed by Rocket-3's expert analysts as a final check.