CN-122021898-A - Sensitive scene-oriented auditable Tibetan language large model system

CN122021898ACN 122021898 ACN122021898 ACN 122021898ACN-122021898-A

Abstract

The invention discloses an auditable Tibetan language large model system facing sensitive scenes, which relates to the technical field of Tibetan language audit models and comprises a compliance safety control module, wherein the compliance safety control module is used for detecting sensitive words and inputting and checking input requests, the compliance safety control module sends checked requests to a retrieval enhancement generation module, the retrieval enhancement generation module injects acquired context information and standard terms provided by a term standardization module into a local large model reasoning module, and the local large model reasoning module returns responses to the compliance safety control module for outputting checking and audit mark. The invention realizes double examination and whole process mark retention of input and output requests by designing the compliance safety control module and the audit system, rapidly identifies and intercepts requests related to sensitive contents such as politics, religions and the like, ensures final content compliance, and improves the reliability and controllability of the system in sensitive scenes.

Inventors

Danzen Rob
Tong Hongjiang
NIMA DUNZHU
WANG PENG
Gesangquni

Assignees

西藏觉罗数字产业管理有限公司

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. The sensitive scene-oriented auditable Tibetan language large model system comprises a compliance safety control module, a search enhancement generation module, a term standardization module and a local large model reasoning module, and is characterized in that the compliance safety control module is used for carrying out sensitive word detection and input examination on an input request, the compliance safety control module sends the examined request to the search enhancement generation module, the search enhancement generation module carries out mixed search on the request to obtain relevant context information, the search enhancement generation module injects the acquired context information and a standard term provided by the term standardization module into the local large model reasoning module, the local large model reasoning module generates Tibetan language response on the injected context information and term, the local large model reasoning module carries out Tibetan language generation through a 70B scale model and a 7B scale model, the local large model reasoning module returns the response to the compliance safety control module to carry out output examination and audit mark, and the system can also carry out offline response.
2. The sensitive scene oriented auditable Tibetan language large model system of claim 1 wherein the local large model reasoning module comprises a 70B scale model and a 7B scale model, the 70B scale model assumes primary answer responsibilities, the 7B scale model takes charge of intention judgment, query rewrite and fast bridging answer tasks, the local large model reasoning module transmits structured rewrite and short instructions preferentially to the 7B scale model, and transmits open question-answer and long context synthesis to the 70B scale model.
3. The sensitive scene oriented auditable Tibetan language large model system of claim 2, wherein the 70B scale model and the 7B scale model are obtained through continuous pre-training, instruction alignment and refusal and compliance bias three-stage training, the continuous pre-training uses Tibetan language single language and high quality alignment corpus to carry out causal language modeling, the instruction alignment uses Tibetan-based question and answer and task instruction samples to carry out supervision fine tuning, the refusal and compliance bias is on strategy labeling data, and legal substitution suggestions are refused and given by a unified template when a strategy is triggered through model learning.
4. The sensitive scene oriented auditable Tibetan large model system of claim 3, wherein the data used for continuous pre-training comprises Tibetan aligned sentence pairs, tibetan monolingual corpus and Tibetan term pairs, and the data is subjected to cleaning, de-duplication, language identification, alignment, segmentation and acceptance flow treatment, the alignment preferably maintains high-confidence aligned pairs, and the segmentation carries out sentence segmentation and paragraph segmentation on long texts.
5. The sensitive scene-oriented auditable Tibetan language large model system is characterized in that the search enhancement generation module adopts a mixed search mode of dictionary matching and vector search, wherein two sets of Aho-Corasick automata in dictionary matching are matched at high speed in use, local FAISS indexes are used for carrying out similarity search in vector search, the scores of the dictionary matching and the vector search are fused in a weighted mode, the dictionary matching weight alpha is 0.7, and the vector search weight beta is 0.3, so that hit precision and semantic correlation are considered.
6. The sensitive scene oriented auditable Tibetan language large model system of claim 5 wherein the retrieval enhancement generation module generates a bridging answer through a 7B scale model, wherein the bridging answer is used as a supplement to matching of a retrieved auxiliary text with a dictionary, and the actual retrieval text is formed by combining a latest multi-round dialogue abstract and the bridging answer and is used for enhancing the coverage of a context memory and a current question.
7. The sensitive scene oriented auditable Tibetan language large model system of claim 1 wherein the term normalization module is executed after retrieval, the term normalization module injects the highest score terms into the context template in a structured manner, and in the generation stage, the model language preference is tilted towards Tibetan language by pre-training and instruction alignment, and language character set mask and post-processing filtering are applied to inhibit generation of non-Tibetan characters and rewrite repair of cross-language paragraphs.
8. The sensitive scene oriented auditable Tibetan language large model system of claim 1 wherein the compliance security control module comprises an input and an output, wherein the input link carries out sensitive word detection, file filtering, personal sensitive information identification and desensitization, and the output link carries out secondary interception and audit system archiving.
9. The sensitive scene oriented auditable Tibetan language large model system of claim 8, wherein the auditing system is arranged in a compliance security control module and is used for recording input and output results, including time stamps, strategy hits and disposal information.
10. The sensitive scene oriented auditable Tibetan language large model system according to claim 1, wherein the system comprises a containerized deployment and an offline delivery, the containerized deployment is supported to run on a local server of a plurality of GPUs, the offline delivery packages images, indexes and data through an offline package catalog, and search resources are preheated when a service is started for stable running and low-delay response in an offline environment, and the search resources comprise a glossary, a FAISS index and an automaton.

Description

Sensitive scene-oriented auditable Tibetan language large model system Technical Field The invention relates to the technical field of Tibetan language audit models, in particular to an auditable Tibetan language large model system oriented to a sensitive scene. Background Along with the rapid development of artificial intelligence technology, a large language model shows strong capability in natural language processing tasks, however, when resources such as Tibetan language are relatively scarce and languages with special application scenes are adopted, the prior art scheme faces a plurality of difficulties, under sensitive scenes such as national culture, regional policies and the like, the prior model lacks effective compliance and safety control mechanisms, the prior model is difficult to accurately identify and filter requests and generation results related to sensitive contents, higher compliance risks exist, and meanwhile, the prior model lacks audit trail functions, so that model behaviors cannot be traced and accountability, and the rigid requirements of data safety and content controllability under the sensitive scenes cannot be met. Patent CN113033180B discloses a service system for automatically generating Tibetan reading problems of primary school, and the above-mentioned patent solves the problems of small number of body forms, low update speed, small amount of manual questions and the like of Tibetan reading teaching materials of primary school. According to the patent, through the designed mixed multi-strategy text screening model, tibetan articles suitable for reading by primary schools can be screened out from large-scale encyclopedic Tibetan texts, and an end-to-end automatic problem generating model is designed, so that the problems of few body forms, low updating speed, small manual problem setting and the like of teaching materials for reading by primary schools are solved, the development of Tibetan teaching in national regions is promoted, and however, optimization space is still reserved in the aspects of compliance and safety control under sensitive scenes related to national culture and the like. Therefore, the application provides the auditable Tibetan large model system which can accurately identify and filter sensitive scenes related to sensitive contents. Disclosure of Invention The invention aims to provide a sensitive scene-oriented auditable Tibetan language large model system so as to solve the technical problem that the conventional Tibetan language model provided in the background technology is difficult to accurately identify and filter sensitive contents. The invention provides a sensitive scene-oriented auditable Tibetan language large model system, which comprises a compliance safety control module, a retrieval enhancement generation module, a term standardization module and a local large model reasoning module, wherein the compliance safety control module is used for carrying out sensitive word detection and input examination on an input request, the compliance safety control module sends the examined request to the retrieval enhancement generation module, the retrieval enhancement generation module carries out mixed retrieval on the request to obtain relevant context information, the retrieval enhancement generation module injects the acquired context information and a standard term provided by the term standardization module into the local large model reasoning module, the local large model reasoning module generates Tibetan language response on the injected context information and term, the local large model reasoning module carries out Tibetan language generation through a 70B scale model and a 7B scale model, the local large model reasoning module returns the response to the compliance safety control module to carry out output examination and audit mark remaining, and the system can also carry out offline response. Preferably, the local large model reasoning module comprises a 70B scale model and a 7B scale model, the 70B scale model bears main answer responsibilities, the 7B scale model is responsible for intention judgment, query rewrite and fast bridging answer tasks, the local large model reasoning module transmits structured rewrite and short instructions to the 7B scale model preferentially, and transmits open question-answer and long context synthesis to the 70B scale model. Preferably, the 70B scale model and the 7B scale model are obtained through continuous pre-training, instruction alignment and refusal and compliance bias three-stage training, the continuous pre-training uses Tibetan single language and high quality alignment corpus to carry out causal language modeling, the instruction alignment uses Tibetan-based question and answer and task instruction samples to carry out supervision fine tuning, the refusal and compliance bias is on strategy labeling data, and legal substitution suggestions are refused and given through a unified template when