Internal testing application is now open! Hithink RoyalFlush Information Network (300033.SZ) asks for financial AI modelHithinkGPT has arrived!

As the first stock of "Internet + Finance + AI" in the A-share market, Hithink RoyalFlush Information Network (300033.SZ) has once again innovated with more than ten years of AI technology accumulation, launching the HithinkGPT large model of Wenzhen. This large model uses the transformer's decoder-only architecture, providing five versions to choose from: 7B, 13B, 30B, 70B, and 130B. It allows for a maximum of 32k text input, supports API interface calls, webpage embedding, co-construction, and private deployment capabilities, and provides users with one-stop annotation and evaluation services. The internal testing application for Wenzhen HithinkGPT large model has now launched, and we sincerely invite you to be one of the first internal testing users to explore the infinite possibilities of AI technology together. I. Condensing Technology - Achieving HithinkGPT HithinkGPT large model brings together the technical essence and data advantages of Hithink RoyalFlush Information Network over the years, aiming to enhance the user and partner experience through AI technology and make investments easier. HithinkGPT has shown outstanding performance in multiple scenarios and holds an absolute leading position in the financial field: General domain: HithinkGPT large model performs comprehensively better than mainstream open-source model Llama-2 on more than ten mainstream benchmark evaluation sets, including C-Eval, GSM8K, MMLU, MATH, and more. Financial domain: Hithink RoyalFlush Information Network has built the HithinkFinEval dataset, which covers 17 financial industry exams, including securities industry exams, fund industry exams, accounting qualification exams, CPA, CFA, and more. In these exams, the Wenzhen HithinkGPT-70B large model achieved an excellent average score of 75.9 and outperformed open-source models in all exam subjects. All these achievements are inseparable from Hithink RoyalFlush Information Network's unique financial data advantages and long-term investment in comprehensive AI technology research and development. We have achieved extreme optimization and technical exploration for large model training and inference, including data coverage, model training, efficient inference, and intelligent security: 1. Comprehensive coverage of financial data: Hithink RoyalFlush Information Network utilizes its own over a decade of accumulated data and publicly available financial data in the market to pre-train financial corpus to the level of trillions of tokens. In addition, we have an automated process for data acquisition, cleaning, and data quality validation, adding thousands of billions of tokens of high-quality pre-training data and hundreds of thousands of pieces of high-quality fine-tuning data every month, ensuring real-time and accurate data. 2. Innovative optimization of model training: We have built a scientific and efficient large model training system, including experimental schemes for data composition, scaling laws, model architecture optimization, distributed training framework optimization, and hardware acceleration technology. This allows us to improve the efficiency of large model training by several times and efficiently complete the training of the HithinkGPT large model family (7B, 13B, 30B, 70B, and 130B) at a lower training cost. 3. Extreme utilization of AI computing power: Hithink RoyalFlush Information Network has built a thousand-card heterogeneous cluster to support model training. For inference, we have applied lossless adaptive layer pruning, communication and underlying operator optimization, as well as multi-data center load balancing, increasing the inference throughput speed of the model by more than eight times. Moreover, we have independently developed large model quantization algorithms, with model accuracy loss below 1% after quantization, and halved the deployment memory requirements, further doubling the inference throughput. 4. Intelligent security protection: We use a lightweight and efficient RLHF (Reinforcement Learning with Human Feedback) solution to align the model's security awareness with human understanding. At the same time, we have independently developed a patch-based large model hotfix technology to quickly respond to and block security vulnerabilities, continuously improving the model's security. We aim to build useful, harmless, and morally correct intelligent partners, ensuring that every user can use them with peace of mind. II. Comprehensive Upgrade - All-round Financial Advisor: Wenzhen Originally loved by investors, investment consultant dialog Siasun Robot&Automation: Hithink RoyalFlush Information Network Wenzhen, has now successfully upgraded based on HithinkGPT as the first intelligent investment advisor product in the domestic financial field to apply large model technology. The large model version of Wenzhen covers 15 business matrices, including A-shares, funds, ETFs, Hong Kong stocks, U.S. stocks, bonds, macroeconomics, etc., with more than 50 types of skills, including queries, analysis, comparisons, interpretations, reasons, predictions, suggestions, backtesting, and more related to the seven investment stages, providing users with comprehensive, accurate, stable, and controllable investment decision support. Compared with the traditional version of Wenzhen, the large model version has the following five advantages: comprehensive real-time financial data, powerful semantic understanding, professional investment advice, vivid expression forms, and controllable content generation, aiming to become the all-round financial advisor for users. 1. More comprehensive real-time data: The large model version of Wenzhen can obtain real-time updates of millions of global financial data indicators and hundreds of thousands of financial-related news generated daily. These data cover various aspects such as stocks, bonds, futures, foreign exchange, commodity prices, macroeconomic indicators, and industry data, ensuring that users always have access to the latest and most comprehensive information. 2. More powerful semantic understanding: Wenzhen has been involved in the field of human-computer interaction in finance more than a decade ago, accumulating billions ofHigh-quality structured financial data plays a crucial role in enhancing semantic understanding and personalized comprehension for users. Additionally, the continuous generation of millions of financial data every day, combined with the self-evolution ability of large models, enables a more accurate understanding of user intentions and needs.3. More Professional Investment Advisory: Based on the user's investment goals, we simulate real-life investment advisory services based on the "user's five KYC tag system, six dimensions of investment analysis, and seven steps of investment process." We provide users with the most scientific and rational investment advice to make investing simpler. 4. More Engaging Expression Formats: Hithink RoyalFlush Information Network's Wencai Da Model has broken the industry's commonly used text format limitations. As of now, we have dynamic line graphs, dynamic dual bar graphs, and 14 other data visualization components, as well as image and video generation tools. These tools can be customized by users and cover 15 core business areas including investment advisory, macroeconomics, information, and investment education. We utilize multimodal technology to make information delivery more efficient and user-friendly. 5. More Controllable Content Generation: In terms of security, we have an embedded intelligent risk identification system that can real-time monitor and recall 31 potential risks from 5 categories of question intentions. Through adversarial training and robustness evaluation, the risk recall rate exceeds 99.5%. We have also invited well-known security institutions and experts in the industry to conduct comprehensive security testing and evaluation of our product, gaining high recognition.