TRS-GPT LLM Platform

TRS-GPT LLM is a professional LLM obtained through industry knowledge enhancement and parameter tuning based on the general large model of the Transformer architecture.

TRS-GPT focuses on industries such as media, finance, network information, public security, patents, and security. It uses its own high-quality data for pre-training. The number of model parameters reaches tens of billions. It supports local privatized deployment and can be deployed on a single A800 or Model tuning is performed on Ascend 910 and can be run on consumer-grade inference cards.

TRS-GPT absorbs the essence of open source big models and conducts independent research and development to meet localization requirements such as independent controllability, safety and compliance.

In response to the current problem of weak Chinese characteristics and professional capabilities of large open source models, TRS-GPT has achieved the Chinese characteristics of the base by expanding the Chinese word list and massive general Chinese corpus, and adapting new Chinese word vectors without interfering with the original model.

Enhance. At the same time, by cleaning and processing data sets in vertical fields such as self-owned media, finance, and government affairs, and conducting professional pre-training enhancements, professional capabilities are improved.

In order to solve the problem that the cost of large model training is too high and the data in the generated content cannot be updated in time, TRS-GPT accesses real-time data through a plug-in knowledge base and combines self-developed vector database technology to realize the integration and utilization of real-time data in professional fields.

To address possible hallucination issues in large models, TRS-GPT aligns content security and values from aspects such as data sources and expert-level knowledge indexing. Key data mainly collects data that can be publicly accessed by mainstream official institutions, such as government departments at all levels, mainstream media, etc. All information disclosed by these institutions strictly follows the "three reviews and three schools" system. All collection sources are manually organized and configured to ensure the "purity" and integrity of the data sources. Using a combination of expert standardized indexing + machine automatic indexing, the collected content information is "refined", including low noise, deduplication, data structuring, data normalization, content labeling, attribute knowledge, Security compliance verification, etc., to align data with mainstream values.

TRS-GPT can effectively reduce the computing resource requirements of large models through pruning, quantification, sparseness, distillation and other deployment optimization solutions. The vertical large model launched by TRS-GPT has tens of billions of parameters, supports local privatized lightweight deployment, can be model tuned on a single A800 or Ascend 910, and can be run on a single consumer-grade inference GPU card.

Product Advantages

Controlled generation technology with enhanced Chinese characteristics

In order to solve the problems of value bias and easy induction of large models, through the injection of Chinese official media and propaganda knowledge, socialist values data construction, value alignment and safety fences, we will provide technical training based on RLHF, DPO and other methods for specific industries or industries. The "values" and "preferences" of the organization, such as national security values, national media values, etc., achieve safe and controllable text generation with enhanced Chinese characteristics.

Credibility verification of generated results

Aiming at the hallucination generation problem common in LLM, the knowledge base search engine technology based on dense vectors is used to fact-check the generated results to detect content inconsistent with the facts, and error revision technology is used to revise the false generated results to ensure The generated text is well-founded, effectively improving the quality of text generation.

Fusion technology of knowledge base and large model

Aiming at the problems of large models such as inability to access external knowledge, untimely knowledge updates, and catastrophic forgetting, a method is proposed that combines incremental fine-tuning with external knowledge base retrieval. The incremental fine-tuning method can regularly update knowledge based on freezing most parameters; the external knowledge base retrieval method can give large models the ability to access external knowledge in real time. Through the combination of the two methods, the effective integration of knowledge base and large model is achieved.

Application platform based on TRS-GPT AgentFlow intelligence

In order to solve the problem that large models cannot use external tools, TRS-GPT AgentFlow platform was developed, which combines the capabilities of knowledge base, tool components and large language models (LLM). It aims to simplify the construction and complex question and answer system through AI Agent technology. Problem handling. Through the intuitive interactive interface, users can quickly introduce the knowledge base, define prompts, and build workflows using drag-and-drop and configuration methods to achieve collaboration between large and small AI models and various AI capabilities without knowing the underlying technical details.
　　

Xinchuang software and hardware support

Support Xinchuang(information technology application innovation)software and hardware products, including OSsuch as Kirin, Tongxin, and Deepin, domestic CPUs such as Phytium, Kunpeng, and Loongson, and GPU products such as Huawei Ascend and Hygon DCU.

Product Functions

TRS-GPT Basic Capabilities

TRS-GPT has ten basic capabilities: content generation, multi-round dialogue, semantic understanding, cross-modal interaction, knowledge-based search, logical reasoning, security compliance, mathematical calculations, programming capabilities and plug-in extensions.

TRS-GPT l has four major innovations: controllable generation technology with enhanced Chinese characteristics, credible verification of generated results by integrating search engines, enhanced cross-modal capabilities by integrating dense vectors, and support for timely updating of external knowledge.

TRS-GPT Dialogue Q&A Platform

The dialogue question and answer platform is a question and answer platform based on the TRS-GPT. The platform has four functional modules: intelligent interaction, statistical analysis, knowledge management and system management.

The intelligent interaction module can output interactive services such as intelligent question and answer, question recommendation, historical records, new dialogue, field selection and other interactive services in the form of text to the user through the dialogue page;

The statistical analysis module has rich data statistics and analysis functions. While recording interactive data, it also supports providing various statistical reports from multiple dimensions such as dialogue, knowledge, users, and public opinion for users to make operational decisions;

The knowledge management module provides domain knowledge management, knowledge graph management, knowledge base management and other functions, allowing users to build relevant knowledge bases according to different scenarios and providing industry knowledge support for Q&A;

The system management module provides unified management and maintenance of user permissions and user logs.

AIGC Intelligent Writing Assistant

AIGC intelligent writing assistant is an official document writing platform based on the TRS-GPT. The platform is designed to help users write various official documents quickly and accurately, and improve the efficiency and quality of official document writing.

The target audience of the product is mainly professionals who need to write official documents frequently, such as government workers, business managers, etc. These users usually need to deal with a large number of tedious official document writing tasks, and the AIGC intelligent writing assistant can provide them with fast and accurate official document writing services and reduce their work burden.

Product features include the following aspects:

Intelligent: The product adopts large model technology, which can automatically identify, analyze and optimize the user's official document writing content, and provide intelligent suggestions and error correction functions, such as spelling check, grammar check, sentence structure optimization, etc.

Simple and easy to use: The user interface and interaction design of the product are simple and easy to use. Users can select the corresponding official document template as needed and fill in the information on the template to automatically generate official documents, reducing the user's learning and operating costs.

Efficiency: The product can quickly and accurately generate various official documents, greatly improving the efficiency and quality of official document writing, allowing users to better cope with tedious official document work.

Media TRS-GPT

Based on its own 100 billion+ Internet public information data, 10 billion+ authoritative source media data, 1 billion government agency data, and 500 million industry information data as professional training data, Torsi has created a large media professional model with main functions covering Content production intelligent assistant, new generation search and recommendation, multi-modal communication and service and other business scenarios.

1. Content production intelligent assistant

The content production intelligent assistant uses AI to provide creative assistance for media practitioners in the content creation process, in order to improve the work efficiency of practitioners, reduce the difficulty of creation, and even optimize the media production process. The service will mainly adopt module privatization deployment or cloud service model. The core functions include intelligent preparation of news titles, article paragraph continuation, intelligent content summary, automatic translation of multi-language articles, article style rewriting, text-based pictures of non-news pictures, etc. Efficient quality improvement assists in completing the following business scenarios:

l Creators can use large models to automatically refine and directional reasoning on topic-related articles to provide material and data basis for content planning. l Creators can easily use prompt words as content topics and use AI to automate article writing. l Creators can give an outline and let AI continue writing and generating in the direction they expect. l The creator lets AI condense the summary or news title based on the writing material. l Creators let AI realize different rewritings of mainstream news and new media styles. l Creators can easily learn from foreign news information through automatic translation and achieve efficient and intelligent dissemination to the world.

2. New generation search and recommendation The new generation search engine built by TRS-GPT has changed the operation mode of traditional search engines, giving it the following advantages: The knowledge-based search engine built by TRS-GPT will strengthen the indexing of past information and the knowledge base and material Library construction and association. Assist customers to realize the following scenarios: l More accurate output results: Large models have stronger language understanding and context awareness capabilities, which can more accurately understand users' search intentions and provide search results that better meet user needs. l A more natural interactive experience: The text generated by the large model is more fluid and coherent, making the search results more readable and giving users a better search experience. l Multi-modal search forms: Large models can not only handle traditional text searches, but also handle search requests in various forms such as images and audio, providing users with more diversified search options. l Combined with the private domain data and knowledge of media users, the large base model and professional knowledge are integrated to provide users with more authoritative and professional knowledge service capabilities. l Expand the collection, retrieval and use of information by the media think tank, and provide customers with more high-quality knowledge services in the future, such as the general secretary’s quotations, policies related to data elements, government public statistics, etc. l Strengthen the ability to intelligently summarize and generate statistics on theme events. Realize knowledge enhancement of real-time media big data. l Achieve pre-training of specified data sets and strengthen machine learning and generation capabilities in different data ranges. l Combined with time series communication data, automatically generate communication summary reports or publicity report summaries, etc.

3. Multi-modal communication and services As one of the core applications of the Metaverse, virtual humans are also an effective carrier for AIGC's customer applications. Torsi uses virtual human services to cover media production and broadcasting, interactive Q&A and other scenarios, mainly reflected in: l Realize the automation of virtual human broadcasting: Use the content generation intelligent assistant to complete the generation of broadcast content (with pictures and texts), and complete the one-click generation of virtual human broadcasting video through the virtual human production and broadcasting platform. Similar applications can also be used for product promotion, etc. Scenes. l Through the pre-training of specific scene data and the use of virtual human interactive question and answer form, the application service coverage of cross-industry virtual and real interaction and other scenarios is realized, and the integration of media in "news + government service business" is deepened.

Financial TRS-GPT

Torsi is based on its own 11 billion+ financial subject data, 10 billion-level industrial indicator data, 3 billion+ industrial element detailed data, 200 million+ industrial dynamic ontology, more than 500+ indexing dimensions, 10,000+ knowledge indexing rules, 100,000+ industry labels are used as professional training data to create a large-scale financial professional model. Its main functions cover business scenarios such as intelligent risk control, intelligent customer service, intelligent investment research, and automatic business batch processing.

1. Risk control and public opinion It provides risk warning assistant, risk report generation assistant, and risk knowledge query assistant functions. Based on the original intelligent risk control products, it comprehensively upgrades the basic technology base and uses large model technology to provide the following functional upgrades: l Acquisition of multi-modal risk information, automatic risk summary, risk classification, risk ranking and information prioritization. l In pre-loan due diligence, post-loan investigation, etc., an intelligent report generation assistant is provided, which can provide automatic functions such as report title generation, catalog generation, and report full text generation, and can integrate large model plug-in data to intelligently generate various risk reports . l Use natural language input to implement intelligent search of massive data for internal data and knowledge data

2. Research report generation Artificial intelligence has obvious advantages in improving the efficiency and scientificity of investment research. Torsi will use large models to launch an investment research intelligent search engine and research report generation assistant. l Expand the sources of investment information through artificial intelligence technologies such as natural language processing and deep learning. The financial information system can capture macroeconomic indicators, public opinion trends, regulatory policies and other data related to investment goals, greatly improving the timeliness of obtaining information, and using natural Language processing technology is used to analyze news articles and comments on social media to better understand market sentiment and trends, and formulate investment strategies based on its prediction results, reducing the workload of investment advisors in financial processing of basic data. l Utilize large-scale model technology capabilities and combine hundreds of macro, meso, and micro related data from Datastar Industry Brain to provide an intelligent research report generation assistant, automatically generate research report titles, research report catalogs, and research report content, and combine various types of Plug-in data provides accurate real-time data insertion and content production.

3. Intelligent customer service (consumer protection) Use large model technology to enable financial institutions to handle complaints immediately and improve customer service capabilities. l Customer service data intelligent analysis assistant: It can complete automated analysis of various types of complaint data such as voice and text to complete core requirements such as customer complaint reasons, complaint channels, complaint appeals, whether it is high-risk, and whether it is repeated. l Customer service complaint handling assistant: Use the complaint knowledge base and complaint handling opinion base to automatically generate complaint handling suggestions, liability determination opinions, troubleshooting opinions, etc., to improve the efficiency of customer service complaints handling and reduce the customer complaint rate.

4. Automatic business batch processing Launched contract approval assistant, consumer protection review assistant, and internal control audit review assistant. l Contract approval assistant: Use large models to reconstruct contract compliance, contract indicator extraction, contract template comparison and other related functions to improve contract approval efficiency. l Consumer protection review assistant: Use large model technology to provide automated review of multi-modal marketing content, product introductions, system specifications and other information, and automatically generate review opinions based on the organization's internal compliance and regulatory systems. l Internal control audit review assistant: Use large model technology to intelligently integrate data sources such as internal audit review information, regulatory regulations, and penalties, and integrate important audit information databases such as the audit results knowledge base and program database to achieve full coverage of audit and inspection information needs and support auditing One-click review of results, knowledge, information and experience, and automatic generation of internal control review opinions.

Government affairs TRS-GPT

Based on its own 1 billion government affairs data, official documents, policy documents, government affairs service guides and other data as professional training data, Torsi has created a large-scale government affairs professional model. Its main functions cover official document auxiliary writing, policy brain and new generation of government affairs interaction. Scenes.

1. Assistance in writing official documents Provide faster and more accurate help and suggestions in the document writing process, support the generation of content suggestions across multiple professional fields, improve article structure, etc., help save the author's time and energy, and improve the efficiency of official document writing. l Based on the official document title and prompt information, the official document summary, table of contents, and table of contents are generated in sequence to form the official document writing result. l Supports automatic generation of coherent and logical official document text based on a small amount of input text. l Supports integration with content editors and serves as an auxiliary writing assistant to interact with the content editing process in real time; it can point out inappropriate vocabulary, correct grammatical errors, improve sentence structure, and provide more vivid and attractive expressions to improve the readability of content Sex and attraction.

2. Policy Analysis l Intelligent interpretation of policies: Supports intelligent interpretation of the latest released policies, with the ability to identify and answer questions using real-time data and access models. l Interpretation of the core content of the policy: Supports the interpretation of the core content of the policy and generates a summary, including overall goals, key tasks, policy measures, etc. l Policy enterprise support: Support the interpretation of enterprise support, subsidies and other information in the policy content, and provide support for enterprise reference. l Industrial policy support: Supports comparison of industrial policies in different regions, or policies in the same region at different time periods, to further provide policy research support for enterprises.

3. New generation of government-citizen interaction It can be used in scenarios such as intelligent Q&A on government websites, real-time consultation on service APPs, and robot guidance in government service halls. Netizens express their needs through natural language, without complicated menus or filters, and can freely express their demands in the most natural and convenient way. l Large language model + multi-modal support, reduce the cost of virtual human creation, and quickly generate one-click synthesis of government-civilian interactive virtual hosts with rich expressions, movements, and forms. l In the interaction between government and citizens, virtual humans are given business support capabilities based on large models, 7*24 online services, multiple rounds of interactions, automatic generation of anthropomorphic reply content, and real-time acquisition of business knowledge. l Better understand and handle regional languages, dialects and cultural backgrounds to provide higher quality customer support and personalized proactive care. l Voice recognition, face recognition, action recognition, and emotion recognition realize multi-modal interaction between real people and virtual people. l Open and easy to use, supporting business customization and local deployment.

Application Scenario

Ping An Bank Audit Assistant

SinoMed Union Intelligent Biomedical Literature Integrated Service Platform

China Daily LLM Project

Event Analysis and Deduction Project

Based on TRS-GPT, Ping An Bank's audit assistant implements services such as regulatory knowledge Q&A, auxiliary generation of audit reports, financial text annotation,text proofreading and error correction, and content recommendation.

The intelligent biomedical literature platform project is supported by The Institute of Medical Information (IMI), Chinese Academy of Medical Sciences and Peking Union Medical College (CAMS & PUMC), aims to create a service platform that is open to the whole country, supports scientific researchers to search for literature intelligently and accurately, and supports users of medical colleges to obtain literature conveniently and efficiently.TRS-GPT provides functions such as structured summary recognition, one-sentence summary generation, and review summary generation.

TRS-GPT provides users with services such as English knowledge QA, Englishnamed entity recognition, term recognition, and reading comprehension multi-choice question generation, etc.

TRS-GPT provides users with event model construction and analysis tools based on deep learning frameworks to help users build Internet event libraries and virtual event libraries, extract relationships between events, and achieve visual analysis; based on technologies such as LLM and knowledge graphs, the machine Intelligent event deduction.