As the first stock of "Internet + Finance + AI" in the A-share market, RoyalFlush has innovated again with more than ten years of artificial intelligence technology accumulation and launched the HithinkGPT model. This large model adopts the decoder-only architecture of Transformer, provides five versions of 7B, 13B, 30B, 70B and 130B, allows up to 32k text input, supports API interface calls, web page embedding, co-construction, privatization deployment and other capabilities, and provides users with one-stop annotation and evaluation services
。 Now that the application for the internal testing of the HithinkGPT large model has been launched, we sincerely invite you to become the first batch of internal testing users to explore the infinite possibilities of AI technology together. Scan or press and hold the QR code below to apply for the internal test.
(Long press the QR code to register)
1.凝练技术-成就HithinkGPT
The HithinkGPT model brings together the technical essence and data advantages of RoyalFlush for many years, and is committed to improving the user experience of users and partners through AI technology, making investment easier. HithinkGPT has excellent performance in multiple scenarios and is in an absolute leading position in the financial field:
通用领域：HithinkGPT大模型在C-Eval、GSM8K、MMLU、MATH
and other more than 10 mainstream benchmark sets, which comprehensively surpassed the mainstream open-source model Llama-2.
Financial field: RoyalFlush has built its own HithinkFinEval dataset, covering 17 financial industry exams such as securities practitioner examination, fund practitioner examination, accountant qualification examination, CPA, CFA, etc. In these exams, the HithinkGPT-70B large model passed the exam with an average score of 75.9 points, and HithinkGPT was better than the open source model in all exam subjects.
The above achievements are inseparable from RoyalFlush's unique data advantages in the financial field and long-term investment in AI comprehensive technology research and development, and we have achieved the ultimate 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: Flush has successfully used its own data accumulation in the past ten years and the financial data published in the market to pre-train the financial corpus to reach trillions of tokens. In addition, we have an automated process for data acquisition, cleansing, and data quality verification, adding hundreds of billions of tokens of high-quality pre-training data and hundreds of thousands of high-quality fine-tuning data every month to ensure the real-time and accuracy of the data.
2. Innovative optimization of model training: We have built a scientific and efficient large model training system, including: data matching experimental scheme, scaling law experimental system, model architecture optimization, distributed training framework optimization, hardware acceleration technology, etc. As a result, our large model training efficiency has been improved several times, and the training of the HithinkGPT large model family (7B, 13B, 30B, 70B and 130B) can be completed efficiently with lower training costs.
3. The ultimate use of AI computing power: Straight Flush has built a kilocalorie heterogeneous cluster to provide support for model training. In terms of inference, we have increased the inference throughput speed of the model by more than eight times by applying lossless adaptive layer selection and pruning, communication and underlying operator optimization, and implementing multi-data center load balancing. In addition, we have independently developed a large model quantization algorithm, which reduces the accuracy loss of the model after quantization by less than 1%, halves the deployment memory requirements, and further increases the throughput of inference by more than two times.
4. Resolute protection of intelligent security: We use a lightweight and efficient RLHF solution to bring the large model's perception of safety on par with that of humans. At the same time, we have created the original patch-based large model hot-repair technology to quickly respond to and plug security vulnerabilities and continuously improve the security of the model. We aim to build a useful, harmless, and correct smart partner to ensure that every user can use it with peace of mind.
2. Comprehensive upgrade - all-round financial advisor: ask for finance
The original investment advisory dialogue robot that was widely loved by investors: Flush Ask Cai, is now successfully upgraded based on HithinkGPT, and is the first robo-advisory product in the domestic financial field to apply large model technology.
The large-scale model version of Qcai covers 15 business matrices such as A-shares, funds, ETFs, Hong Kong stocks, U.S. stocks, bonds, and macro, and more than 50 types of skills, including inquiry, analysis, comparison, interpretation, reasons, predictions, suggestions, and backtesting of seven major investment links, providing users with comprehensive, accurate, stable and controllable investment decision-making support.
Compared with the traditional model, the large-scale model version of Ask Cai has the following five characteristic advantages: comprehensive real-time financial data, strong semantic understanding, professional investment advisory suggestions, vivid expression forms, and controllable content generation, and is committed to becoming an all-round financial advisor for users:
- More comprehensive real-time data: The model can obtain millions of real-time updated financial data indicators around the world and hundreds of thousands of financial-related information generated every day. This data covers stocks, bonds, futures, foreign exchange, commodity prices, macroeconomic indicators, industry data and other aspects to ensure that users can always get the latest and most complete information.
- More powerful semantic understanding: Wencai has been involved in the field of financial human-computer interaction more than ten years ago, and has accumulated billions of complete, professional and high-quality structured financial data, which has inherent advantages in improving semantic understanding and user's personalized understanding ability.
- More professional investment advisory advice: Focusing on the user's investment goals, based on the "user's five KYC label systems, six dimensions of investment analysis, and seven steps of the investment process", it simulates the logic of real investment advisory services, gives users the most scientific and reasonable investment advice, and makes investment easier.
-
Up to now, we have 14 kinds of data visualization components such as dynamic line charts and dynamic double-column charts, as well as two rich media visualization tools for picture generation and video generation, and support user customization, covering 15 core business areas such as investment advice, macro, information, investment education, etc., using multimodal technology to make information transmission more efficient and humane.
5. More controllable content generation: In terms of security, through the built-in intelligent risk identification system, it can monitor and recall the question intentions of 31 potential risks in 5 categories in real time, and the risk recall rate exceeds 99.5% through systematic adversarial training and robust evaluation.
Source: Straight Flush
Disclaimer: This article is a reprint and does not represent the position of this journal and does not constitute investment advice.