Data processing method, device and program product

Abstract

The invention provides a data processing method and device and a program product, and relates to the technical field of computers. The method comprises the following steps: acquiring a data analysis request, wherein the data analysis request is used for requesting to analyze data specified by a user and generating a data analysis report; a pre-trained big language model is applied to analyze data specified by a user, an initial analysis link tree, a data analysis report draft and a next analysis plan are generated and displayed based on the diffusion technology, and the big language model is constructed and trained based on the diffusion technology; and executing a next-step analysis plan by applying a large language model to update the displayed initial analysis link tree, the data analysis report draft and the next-step analysis plan until a final analysis link tree and a final data analysis report are obtained. According to the embodiment of the invention, the overall analysis speed can be accelerated, and data analysis report draft preview can be realized.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of computer technology, and more particularly, to a data processing method, device and program product. BACKGROUND

[0002] In the current technical environment, large language models (LLMs) have been widely applied to deep analysis of data and automatic generation of detailed data analysis reports. These models can process and understand large amounts of unstructured information and extract valuable insights from them to support decision-making. However, despite the strong capabilities demonstrated by existing LLM technology, they still face several challenges when used to generate data analysis reports, such as long serial analysis links in the analysis report generation process and the inability to implement draft previews.

[0003] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0004] The present disclosure provides a data processing method, device and program product, which at least partially solves the problem of long serial analysis links in the analysis report generation process and the inability to implement draft previews of large language models (LLMs).

[0005] Other characteristics and advantages of the present disclosure will become apparent from the following detailed description, or will be learned by practice of the present disclosure.

[0006] According to one aspect of the present disclosure, a data processing method is provided, comprising: obtaining a data analysis request, the data analysis request being used to request analysis of user-specified data and generation of a data analysis report; applying a pre-trained large language model to analyze the user-specified data, and generating and displaying an initial analysis link tree, a data analysis report draft and a next step analysis plan based on a diffusion technique, the large language model being constructed and trained based on the diffusion technique; applying the large language model to perform the next step analysis plan to update the displayed initial analysis link tree, data analysis report draft and next step analysis plan until a final analysis link tree and data analysis report are obtained.

[0007] In one embodiment of this disclosure, before applying a pre-trained large language model to analyze user-specified data, the method includes: acquiring historical data from the process of generating a data analysis report, the historical data including system prompt words, user data analysis requests, tools or services used in the analysis, and the current analysis report, which may be a draft or a final data analysis report; masking some lexical units in the current analysis report to obtain a masked analysis report; constructing diffusion training data based on the system prompt words, user data analysis requests, tools or services used in the analysis, and the masked analysis report; and training the large language model to be trained using the diffusion training data to obtain the trained large language model.

[0008] In one embodiment of this disclosure, the historical data for the data analysis report generation process originates from the process of a third-party large language model generating the data analysis report. This third-party large language model is not built and trained based on diffusion technology.

[0009] In one embodiment of this disclosure, a portion of the words in the current analysis report are masked to obtain a masked version of the analysis report, including: generating a masking rate; and based on the masking rate, masking the random words in the current analysis report that are proportional to the masking rate to obtain a masked version of the analysis report.

[0010] In one embodiment of this disclosure, multiple mask rates are generated for the current analysis report of the same historical data, and each mask rate corresponds to a set of diffusion training data.

[0011] In one embodiment of this disclosure, training a large language model to be trained using diffusion training data to obtain a trained large language model includes: training the large language model to be trained using diffusion training data and performing supervised fine-tuning on the large language model to obtain a pre-trained large language model.

[0012] In one embodiment of this disclosure, the application of a large language model to execute a next analysis plan to update the displayed initial analysis link tree, data analysis report draft, and next analysis plan includes: calling a data processing tool to obtain the data required for the next analysis plan, the data processing tool including a data retrieval tool and / or a search tool; adding the obtained data required for the next analysis plan to the current analysis link, so that the large language model can perform analysis based on the updated information to update the displayed initial analysis link tree, data analysis report draft, and next analysis plan.

[0013] In one embodiment of this disclosure, updating the initial analysis link tree, the draft data analysis report, and the next analysis plan includes: performing analysis by combining the data required for the next analysis plan and the user-specified data, and continuing to edit the draft data analysis report based on diffusion technology; analyzing and planning the analysis to be performed and the data to be acquired in the next analysis plan based on the data required for the next analysis plan and the current draft data analysis report; and adding new branches to the initial analysis link tree based on the analysis required for the next analysis plan.

[0014] In one embodiment of this disclosure, when all necessary data analysis is completed and there are no new branches to explore in the current analysis link tree, the currently displayed analysis link tree and data analysis report draft are used as the final analysis link tree and data analysis report.

[0015] According to another aspect of this disclosure, an electronic device is provided, comprising: a memory for storing instructions; and a processor for calling the instructions stored in the memory to implement the above-described data processing method.

[0016] According to another aspect of this disclosure, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the data processing method described above.

[0017] According to another aspect of this disclosure, a computer program product is provided, which stores instructions that, when executed by a computer, cause the computer to perform the data processing method described above.

[0018] According to another aspect of this disclosure, a chip is provided, including at least one processor and an interface; the interface is used to provide program instructions or data to the at least one processor; the at least one processor is used to execute the program instructions to implement the above-described data processing method.

[0019] The data processing method, device, and program products provided in this disclosure utilize a large language model built and trained based on diffusion technology. Diffusion technology allows the model to process information in parallel at each step, thereby accelerating the overall analysis speed, significantly shortening the serial analysis chain, and reducing the computational resource requirements of context length. The function of displaying the initial analysis chain tree, the draft data analysis report, and the next analysis plan greatly enhances user interactivity and transparency. Users can view the current data analysis report at any time during the analysis process. As each step of the analysis plan is executed, the system automatically updates the displayed content (i.e., the analysis chain tree and the draft data analysis report) until all necessary analysis steps are completed. This iterative update mechanism ensures that each iteration is based on the latest data, guaranteeing the accuracy and timeliness of the results. Furthermore, from the initial analysis overview to the detailed final report, the user remains informed and in control, enhancing the user experience.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0022] Obviously, the accompanying drawings described below are merely some embodiments of this disclosure. Those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0023] Figure 1 This diagram illustrates a data processing method according to an embodiment of the present disclosure.

[0024] Figure 2a , Figure 2b and Figure 2c This diagram illustrates the process of content change in the output interface of the large language model in this embodiment of the present disclosure.

[0025] Figure 3 This diagram illustrates the display content of a large language model output interface according to an embodiment of the present disclosure.

[0026] Figure 4 This diagram illustrates the model training flowchart in an embodiment of the present disclosure.

[0027] Figure 5 This diagram illustrates a mask processing flowchart in an embodiment of the present disclosure.

[0028] Figure 6 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0029] To facilitate understanding of the technical solutions of this disclosure, the disclosure will be further described below with reference to the accompanying drawings.

[0030] The terms "first" and "second," etc., in this disclosure, claim, and drawings are used only to distinguish different objects and not to describe a particular order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0031] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this disclosure. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] In this disclosure, "at least one (item)" refers to one or more, "more than" refers to two or more, "at least two (items)" refers to two or three or more, and "and / or" is used to describe the relationship between related objects, indicating that there can be three relationships. For example, "A and / or B" can indicate three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. "Or" indicates that there can be two relationships, such as only A exists or only B exists; when A and B are not mutually exclusive, it can also indicate three relationships, such as only A exists, only B exists, or both A and B exist simultaneously. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items. For example, at least one (item) of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c".

[0033] An LLM Agent is an LLM-driven workflow. To complete a (complex) task, the LLM agent plans and executes appropriate steps on its own. This process may involve calling external tools. The LLM receives the results returned by the tools and then continues to execute subsequent steps until the task is completed.

[0034] In the current scenario of generating analysis reports, the execution flow of an LLM agent is generally as follows:

[0035] Input: Analyze query (e.g., help me analyze the changes in GTV metrics, both month-on-month and year-on-year, and the reasons behind them).

[0036] Output:

[0037] Step 1: Thinking (e.g., first analyze the overall situation and business lines) -> Call the data retrieval tool (e.g., query the overall and business line data);

[0038] Step 2: Thinking (e.g., discovering that business lines A and B are performing outstandingly / poorly, deciding to further analyze and explore the reasons) -> calling data retrieval tools (e.g., querying data for the sub-categories of business lines A and B);

[0039] Step 3: Thinking (e.g., discovering that the poor performance of category a1 led to the poor performance of business A, while the outstanding performance of categories b3 and b4 led to the good performance of business B. Next, analyze a1, b3, and b4 for a more in-depth investigation) -> call the data retrieval tool;

[0040] ...

[0041] Step n: Thinking (e.g., I have obtained enough data and performed sufficient interpretation; now I will begin writing the final summary report) -> Final Report.

[0042] The above solution has the following problems:

[0043] Excessive context length: Each step of this process is accumulated and stored in the multi-turn dialogue context of the large model. As the number of steps increases, the length of the stored interaction history increases, and the computational cost of each step increases with the number of steps. The significant increase in context length also poses a challenge to the performance and effectiveness of the model.

[0044] Lack of draft preview function: In addition, the current analysis process paradigm (continuously calling tools -> finally writing a report) lacks the interpretability of intermediate steps and cannot preview the draft report during the analysis process (while human logic can make the process of thinking and calling tools while writing a draft report).

[0045] The deficiencies of the above solutions and the proposed solutions are the result of the inventor's practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed in this disclosure below should be considered as the inventor's contribution to this disclosure.

[0046] It is understood that the data involved in this disclosure (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and provisions. Before using the technical solutions disclosed in the embodiments of this disclosure, users shall be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and their authorization shall be obtained.

[0047] The following detailed description of this exemplary implementation method is provided in conjunction with the accompanying drawings and embodiments.

[0048] Figure 1 A flowchart of a data processing method according to an embodiment of this disclosure is shown, such as Figure 1 As shown, the data processing method provided in this embodiment includes S101-S104.

[0049] In S101, a data analysis request is obtained. The data analysis request is used to request the analysis of user-specified data and the generation of a data analysis report.

[0050] In S102, a pre-trained large language model is used to analyze the user-specified data.

[0051] The aforementioned large language model is built and trained based on diffusion techniques; that is, it is a diffusion model. Diffusion models belong to generative models, which generate data by progressively adding noise to the data and then learning how to remove the noise in reverse. The aforementioned large language model uses diffusion techniques to generate the current analysis results, namely the initial analysis link tree, a draft data analysis report, and the next analysis plan.

[0052] Traditional large language models for text generation are autoregressive models. These models predict the next word sequentially based on the generated word sequence, generating only one word or token at a time. This sequential process is slow and limits the quality and consistency of the output. Unlike most previous language models that predict text generation "from left to right," the large language model of this disclosure learns to generate output by progressively optimizing noise, which can significantly improve generation speed and reduce training uncertainty.

[0053] In S103, the current analysis results of S102 are displayed in real time, including the initial analysis link tree, the draft data analysis report, and the next analysis plan.

[0054] Before the stopping condition is met, the large language model is used to execute the next analysis plan to update the initial analysis link tree, the draft data analysis report, and the next analysis plan until the stopping condition is met, resulting in the final analysis link tree and data analysis report, i.e., S104.

[0055] In some embodiments, obtaining the final analysis link tree and data analysis report upon reaching the stopping condition can be achieved by using the currently displayed draft analysis link tree and data analysis report as the final analysis link tree and data analysis report when all necessary data analysis is completed and the current analysis link tree no longer has new branches to explore.

[0056] The data processing method based on diffusion technology disclosed in this embodiment significantly reduces the long context of the serial analysis link, which not only improves the efficiency and accuracy of data analysis report generation, but also enables real-time preview of report drafts, greatly enhancing the user experience and allowing even non-professional users to easily participate in the data analysis process.

[0057] In some embodiments, updating the displayed initial analysis link tree, data analysis report draft, and next step analysis plan includes: performing analysis by combining the data required for the next step analysis plan with user-specified data, and continuing to edit the data analysis report draft based on diffusion technology; analyzing and planning the analysis to be performed and the data to be acquired in the next step analysis plan based on the data required for the next step analysis plan and the current data analysis report draft; and adding new branches to the initial analysis link tree based on the analysis required for the next step analysis plan.

[0058] The aforementioned process of further editing the draft data analysis report based on diffusion technology utilizes the diffusion characteristics of the large language model in this disclosure embodiment. It doesn't simply add content, but rather treats the entire draft data analysis report as a "structure to be optimized." The large language model can detect whether new data (i.e., data required for the next analysis plan) supports or refutes existing conclusions in the draft. If it supports existing conclusions, no processing is needed; if it refutes existing conclusions, the relevant content in the original draft data analysis report needs to be deleted. Furthermore, the large language model also generates new analysis paragraphs, chart descriptions, or visualization suggestions based on the new data. In other words, the above editing process can include modifying, deleting, or reorganizing existing content to ensure logical coherence and accurate conclusions. The diffusion process allows the model to gradually "de-noise" from "noise" (inconsistent or outdated information), generating a more accurate and complete report version.

[0059] Based on the data required for the next analysis plan and the current draft data analysis report, the analysis and planning of the next analysis plan includes the analysis to be performed and the data to be acquired. This can be achieved by using a large language model to analyze the missing content in the current draft data analysis report, and further using the existing data to analyze whether new data needs to be retrieved to complete the missing content, and what data is needed, thereby obtaining the analysis to be performed and the data to be acquired in the next analysis plan.

[0060] In this embodiment of the disclosure, the analysis link tree, draft data analysis report, and next step analysis plan output by the large language model are continuously updated and changed. This change process can be as follows: Figure 2a , Figure 2b and Figure 2c As shown, the analysis link tree and data analysis report are constantly updated and changed based on the original version. Users can directly see the current version of the data analysis report, resulting in a better experience.

[0061] In some embodiments, the process of continuously updating and changing the analysis link tree, draft data analysis report, and next analysis plan output by the large language model displayed on the interface may include markers indicating the editing and modification process of the analysis link tree, draft data analysis report, and next analysis plan. As a specific example, the analysis link tree, draft data analysis report, and next analysis plan output by the large language model may be as follows: Figure 3 As shown.

[0062] This disclosure utilizes a diffusion language model to achieve parallel generation of the thought process and the report (draft). It significantly reduces the context length during step iterations, and the more complex the report, the greater the improvement in report generation speed. Furthermore, it provides a draft preview for the report generation process, allowing users to preview the report at each step without having to wait for a long time.

[0063] In some embodiments, applying a large language model to execute the next analysis plan to update the displayed initial analysis chain tree, data analysis report draft, and next analysis plan may include: calling data processing tools to obtain the data required for the next analysis plan, the data processing tools including data retrieval tools and / or search tools; adding the obtained data required for the next analysis plan to the current analysis chain, so that the large language model can perform analysis based on the updated information to update the displayed initial analysis chain tree, data analysis report draft, and next analysis plan.

[0064] In this embodiment of the disclosure, new data can be introduced by calling data processing tools each time the next analysis plan is executed, thereby improving the content of the data analysis report.

[0065] In some embodiments, a large language model can be built and trained before applying a pre-trained large language model to analyze user-specified data. The training process of the large language model can be as follows: Figure 4 As shown, it includes S401-S404.

[0066] In S401, historical data of the data analysis report generation process is obtained. Historical data includes system prompts, user data analysis requests, tools or services used in the analysis, and the current analysis report, which may be a draft or the final data analysis report.

[0067] In some embodiments, the historical data for the data analysis report generation process comes from the process of a third-party large language model generating the data analysis report. This third-party large language model is not built and trained based on diffusion technology.

[0068] In some embodiments, the historical data used in the process of generating data analysis reports may also be manually written.

[0069] This disclosure provides rich material for subsequent training by collecting historical data (including system prompts, user requests, tools or services used, and current analysis reports). This includes not only data generated by third-party large language models built using non-diffusion techniques, but also high-quality samples written by humans, increasing data diversity and model generalization ability.

[0070] In S402, some words in the current analysis report are masked to obtain a masked version of the analysis report.

[0071] By randomly masking a portion of the terms in the current analysis report, a masked version of the analysis report is created. This process simulates the reverse recovery mechanism of the diffusion model, namely the ability to recover the complete text from partial information, which helps to improve the robustness and accuracy of the model in practical applications.

[0072] In S403, diffusion training data is constructed based on system prompts, user data analysis requests, tools or services used in the analysis, and masked analysis reports.

[0073] By constructing a diffusion training dataset based on the above elements, the model can learn in an environment containing the complexity of real-world scenarios. This approach effectively solves the problem of excessively long traditional serial analysis links and reduces the complexity of context management.

[0074] In S404, the large language model to be trained is trained using diffusion training data to obtain the trained large language model.

[0075] Using this data, the large language model was fine-tuned to obtain a model specifically optimized for generating data analysis reports. This model can handle data analysis tasks more efficiently, provide real-time draft previews, and dynamically adjust the analysis path based on feedback, significantly improving the speed and quality of report generation.

[0076] Existing large language model (LLM agent) generally adopts a serial analysis chain (thinking → data retrieval → rethinking → re-retrieval of data...). The interaction history of each step (thinking, tool call, returned data) accumulates continuously, resulting in a surge in context length, high computational cost, response delay, and users cannot preview any content before the final report is generated.

[0077] This disclosed embodiment utilizes a dedicated large language model built on diffusion technology to achieve a paradigm shift: after receiving all current data at each step, the model directly performs parallel and iterative editing of the "report draft." It eliminates the need to repeatedly store lengthy analysis histories in the context; only the current data and a partially masked draft are required, significantly reducing computational overhead and latency. The result generated at each diffusion step is the displayable report draft, allowing users to view the analysis progress in real time, receive immediate feedback, and experience a process closer to that of a human analyst "analyzing and writing simultaneously." This solution is not simply an application of diffusion technology, but rather a reconstruction of it into a novel, efficient, low-latency, and interactive generative architecture for complex analysis tasks, achieving a qualitative leap in generation efficiency, user experience, and system performance.

[0078] In some embodiments, such as Figure 5 As shown, some words in the current analysis report are masked to obtain a masked version of the analysis report, including S501-S502.

[0079] In S501, the mask rate is generated;

[0080] In S502, based on the masking rate, the random words in the current analysis report that are proportional to the masking rate are masked to obtain a masked version of the analysis report.

[0081] In some embodiments, multiple mask rates are generated for the current analysis report of the same historical data, with each mask rate corresponding to a set of diffusion training data. That is, by using a set of diffusion training data as a training sample, different training samples can be obtained by processing the same historical data with different mask rates.

[0082] Understandably, the above scheme masks the "random words" in the current analysis report, which constitute a proportion of the masking rate. Therefore, theoretically, "infinitely many" random masks can be applied to the current analysis report; more accurately, there are an exponential number of different masking methods (i.e., multiple permutations and combinations). This is a key and powerful feature in training diffusion language models.

[0083] Assuming the current analysis report has N tokens, if N = 100 and the random masking rate is 0.3 (meaning 30% of the tokens are randomly masked), then when applying the random mask, 30 positions need to be randomly selected from these N positions for masking. The total number of possible masking combinations is approximately 3 × 10^3. 25 While not "infinite" in the mathematical sense, the number of possible mask permutations and combinations is extremely large in practical applications, almost "inexhaustible." When training a large language model, it does not memorize a fixed "masked position," but learns to dynamically infer the content of any position based on the context, thereby enhancing the robustness of the model.

[0084] Furthermore, the large number of mask combinations mentioned above makes the diffusion process not a single path, but a probabilistic process. This is equivalent to covering a large number of possible noise-adding paths during training, which helps the model learn the inverse denoising distribution better.

[0085] This disclosure enhances the model's ability to understand and generate high-quality data analysis reports by introducing diffusion techniques and a meticulous data preparation process, while also improving user experience and work efficiency.

[0086] In some embodiments, applying diffusion training data to train a large language model to be trained to obtain a trained large language model includes: applying diffusion training data to train the large language model to be trained and performing supervised fine-tuning on the large language model to obtain a pre-trained large language model.

[0087] In summary, this disclosure embodiment constructs and trains a large language model based on diffusion technology, and then applies the large language model to analyze data and generate a data analysis report. The large language model and its training process of this disclosure embodiment are described in detail below.

[0088] During the pre-training phase, the model randomly masks all tokens (also called words) in the input text. In the SFT phase, only the response portion to be generated is masked. The entire diffusion output process is a gradual process from full masking to complete demasking. In each step, the model can simultaneously predict all masked locations and flexibly remask.

[0089] Analysis of the generation paradigms of the analysis link (indicated by token):

[0090] Autoregressive analysis chain generation: prompt_tokens + user_tokens + step_1_tokens + step_2_tokens + ... + step_n_tokens + report_tokens;

[0091] Among them, step_k_tokens contains the thought process for the current step, the request to call the tool, and the data returned after the tool is successfully called.

[0092] Diffusion-based analysis chain generation: Each step is: prompt_tokens + user_tokens + tool_tokens + diffusion_tokens_for_each_step;

[0093] The tool_tokens contains the accumulated data returned by the tools.

[0094] In this embodiment of the disclosure, the model input includes system prompts, data analysis requests (also known as user input), and initial diffusion content.

[0095] System prompt: Clearly define the current task background and requirements.

[0096] User input: Analysis requirements (e.g.: Help me analyze the changes in user traffic on Meituan Waimai in Q1).

[0097] Initial diffusion content:

[0098] When the first step of analysis is performed, the initial diffusion content is empty, which means that the model needs to go through the process of going from a complete mask to a complete demask.

[0099] Starting from the second step of analysis, the initial diffusion content is the diffusion generation result of the previous step, which means that the model needs to go through the process of going from partial masking to complete demasking (which can be understood as directly editing the draft of the previous step: some content will be retained, and other content will be added, deleted or modified).

[0100] In this embodiment of the disclosure, the model output is: at each step, a diffusion output is performed, that is, the model performs a progressive process from full masking (first step analysis) or partial masking (starting from the second step analysis) to complete demasking based on the input.

[0101] For the data analysis scenarios described above, the output includes:

[0102] Analyzing the link tree helps the model understand and track the analytical approach;

[0103] From the initial draft of a data analysis report, to the final step, it becomes the official report;

[0104] The next step of the analysis plan guides what data to acquire in the next stage of the analysis, including the necessary query requirements.

[0105] The overall performance of the analysis link in this disclosure embodiment is an iterative, editable report writing process, with similar examples.Figures 2a-2c .

[0106] In this embodiment of the disclosure, based on the model output, the corresponding tool (such as a data retrieval tool or a search tool) can be called, and the returned data will be added to tool_tokens.

[0107] When constructing training data for a model, progressive, iteratively editable data can be created through manual writing or by calling other LLMs, and then a random mask can be applied to it. Specifically, assuming we have the input (system_prompt, user_query, tools) and output (draft_report) during the iteration process, i.e., (system_prompt, user_query, tools, draft_report), then to construct the training data for the diffusion language model, a random mask needs to be applied to draft_report. The process of applying the random mask is as follows:

[0108] A) Generate a random masking rate theta (a probability between 0 and 1): The larger theta is, the more tokens are masked, which means the early stage of the diffusion process; the smaller theta is, the fewer tokens are masked, which means the late stage of the diffusion process (meaning that the diffusion generation process is only a few tokens away from completion).

[0109] B) Based on the theta generated by A, the random tokens in the response with a proportion of theta are transformed into a [mask] mask, thus obtaining draft_report_masked.

[0110] C) Diffusion training data was obtained:

[0111] The input is: (system_prompt,user_query,tools,draft_report_masked);

[0112] The output is: (system_prompt,user_query,tools,draft_report).

[0113] The model only learns draft_report_masked and draft_report; the prompt, query, and tool parts remain unchanged and are not involved in the learning process.

[0114] An analysis pipeline contains multiple analysis steps, each corresponding to a data pair (system_prompt, user_query, tools, draft_report). Starting from each data pair, countless random masks (multiple permutations and combinations of random masks) can be applied to construct a massive amount of training data.

[0115] Similar to mainstream LLM models, after obtaining training data, SFT fine-tuning is performed on the pre-trained diffuse large language model to finally obtain the fine-tuned diffuse large language model.

[0116] In this embodiment, the query statement for each step does not need to be included in the analysis chain; it only needs to be generated at each step. Furthermore, the analysis of the large model at each step does not require generating a thought process, which can significantly reduce the long context of the serial analysis chain. The tokens of the report content in the diffusion generation process are generated in parallel, which is faster than autoregressive generation and can improve generation efficiency. The diffusion generation results at each step can be displayed to the user in real time, while the autoregressive generation paradigm cannot display to the user until the final step.

[0117] The following reference Figure 6 This describes the electronic device provided in the embodiments of this disclosure. Figure 6 The electronic device 600 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0118] Figure 6 This diagram illustrates the architecture of an electronic device 600 provided in an embodiment of the present invention. Figure 6 As shown, the electronic device 600 includes, but is not limited to, at least one processor 610 and at least one memory 620. The memory 620 is used to store instructions.

[0119] In some embodiments, memory 620 may include a readable medium in the form of volatile memory cells, such as random access memory (RAM) 6201 and / or cache 6202, and may further include read-only memory (ROM) 6203.

[0120] In some embodiments, the memory 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0121] In some embodiments, memory 620 may store an operating system. This operating system may be a real-time operating system (RTX), such as Linux, UNIX, Windows, or OS X.

[0122] In some embodiments, the memory 620 may also store data.

[0123] As an example, processor 610 can read data stored in memory 620, which may be stored at the same memory address as the instructions, or the data may be stored at a different memory address than the instructions.

[0124] Processor 610 is configured to invoke instructions stored in memory 620 to implement the steps described in the "Exemplary Methods" section above, according to various exemplary embodiments of this disclosure. For example, processor 610 can execute the steps of the above-described data processing method embodiments.

[0125] It should be noted that the processor 610 described above can be a general-purpose processor or a special-purpose processor. The processor 610 may include one or more processing cores, and the processor 610 executes various functional applications and data processing by running instructions.

[0126] In some embodiments, processor 610 may include a central processing unit (CPU) and / or a baseband processor.

[0127] In some embodiments, the processor 610 may determine an instruction based on the priority identifier and / or function category information carried in each control instruction.

[0128] In this disclosure, the processor 610 and memory 620 can be configured separately or integrated together. As an example, the processor 610 and memory 620 can be integrated on a single board or a system-on-chip (SOC).

[0129] like Figure 6 As shown, electronic device 600 is embodied in the form of a general-purpose computing device. Electronic device 600 may also include bus 630.

[0130] Bus 630 can represent one or more of several types of bus structures, including a memory bus or memory controller, peripheral bus, graphics acceleration port, processor, or a local bus using any of the various bus structures.

[0131] Electronic device 600 can also communicate with one or more external devices 640 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem, etc.). Such communication can be performed through input / output (I / O) interface 650.

[0132] Furthermore, the electronic device 600 can also communicate with one or more networks (such as local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via the network adapter 660.

[0133] like Figure 6 As shown, network adapter 660 communicates with other modules of electronic device 600 via bus 630.

[0134] It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0135] It is understood that the structures illustrated in the embodiments of this disclosure do not constitute a specific limitation on the electronic device 600. In other embodiments of this disclosure, the electronic device 600 may include more than Figure 6 This may involve more or fewer components, or combining certain components, or splitting certain components, or different component arrangements. Figure 6 The components shown can be implemented in hardware, software, or a combination of both.

[0136] This disclosure also provides a computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the data processing method described in the above method embodiments.

[0137] In this disclosure, the computer-readable storage medium is one capable of sending, propagating, or transmitting computer instructions for use by or in connection with an instruction execution system, apparatus, or device. As an example, the computer-readable storage medium is a non-volatile storage medium.

[0138] In some embodiments, more specific examples of computer-readable storage media in this disclosure may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, USB flash drives, portable hard drives, or any suitable combination of the foregoing.

[0139] In this embodiment of the disclosure, the computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, wherein computer instructions (readable program code) are carried.

[0140] In some examples, computational instructions contained on a computer-readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0141] This disclosure also provides a computer program product storing instructions that, when executed by a computer, cause the computer to perform the data processing method described in the above-described method embodiments. These instructions may be program code. In specific implementations, the program code may be written using any combination of one or more programming languages. The program code may execute entirely on a user's computing device, partially on a user's device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0142] This disclosure also provides a chip, including at least one processor and an interface; the interface is used to provide program instructions or data to at least one processor; the at least one processor is used to execute the program instructions to implement the data processing method described in the above method embodiments.

[0143] In some embodiments, the chip may further include a memory for storing program instructions and data, the memory being located within or outside the processor.

[0144] Those skilled in the art will understand that all or part of the steps of the above embodiments can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, which can be collectively referred to as "circuit", "module" or "system".

[0145] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein.

[0146] This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A data processing method, characterized in that, include: Obtain a data analysis request, which is used to request the analysis of user-specified data and the generation of a data analysis report; The pre-trained large language model is used to analyze the user-specified data, and an initial analysis link tree, a draft data analysis report, and a next step analysis plan are generated and displayed based on diffusion technology. The large language model is built and trained based on diffusion technology. The large language model is applied to execute the next analysis plan to update the initial analysis link tree, the draft data analysis report, and the next analysis plan until the final analysis link tree and data analysis report are obtained.

2. The method according to claim 1, characterized in that, Before the application of the pre-trained large language model to analyze the user-specified data, the method includes: The process of generating data analysis reports is to acquire historical data, including system prompts, user data analysis requests, tools or services used in the analysis, and the current analysis report, which may be a draft or the final data analysis report. The historical data of the data analysis report generation process originates from the process of generating data analysis reports using a third-party large language model, which is not built and trained based on diffusion technology. Mask some of the terms in the current analysis report to obtain a masked version of the analysis report; Based on the system prompts, the user's data analysis request, the tools or services used in the analysis, and the masked analysis report, diffusion training data is constructed. The large language model to be trained is trained using the diffusion training data to obtain the trained large language model.

3. The method according to claim 2, characterized in that, The step of masking a portion of the terms in the current analysis report to obtain a masked version of the analysis report includes: Generate mask rate; Based on the masking rate, the random words in the current analysis report that are proportional to the masking rate are masked to obtain a masked version of the analysis report.

4. The method according to claim 3, characterized in that, For the current analysis report of the same historical data, multiple mask rates are generated, and each mask rate corresponds to a set of diffusion training data.

5. The method according to claim 2, characterized in that, The process of using the diffusion training data to train the large language model to be trained, and obtaining the trained large language model, includes: The large language model to be trained is trained using the diffusion training data, and the large language model is then fine-tuned under supervision to obtain the pre-trained large language model.

6. The method according to claim 1, characterized in that, The application of the large language model to execute the next analysis plan, updating the displayed initial analysis link tree, the draft data analysis report, and the next analysis plan, includes: The data processing tools are invoked to obtain the data required for the next step of the analysis plan. The data processing tools include data retrieval tools and / or search tools. The data required for the next step analysis plan is added to the current analysis chain so that the large language model can perform analysis based on the updated information, thereby updating the initial analysis chain tree, the draft data analysis report, and the next step analysis plan.

7. The method according to claim 6, characterized in that, The updated display of the initial analysis chain tree, the draft data analysis report, and the next analysis plan includes: The data required for the next step of the analysis plan and the user-specified data are analyzed together, and the data analysis report draft from the previous step is further edited based on diffusion technology. Based on the data required for the next step of the analysis plan and the current draft data analysis report, analyze and plan the analysis to be performed and the data to be obtained in the next step of the analysis plan; Based on the analysis required in the next step of the analysis plan, new branches are added to the initial analysis link tree.

8. The method according to claim 1, characterized in that, When all necessary data analysis is completed and there are no new branches to explore in the current analysis tree, the currently displayed analysis tree and data analysis report draft will be used as the final analysis tree and data analysis report.

9. An electronic device, characterized in that, include: Memory, used to store instructions; A processor is configured to invoke instructions stored in the memory to implement the data processing method as described in any one of claims 1-8.

10. A computer program product, characterized in that, The computer program product stores instructions that, when executed by a computer, cause the computer to perform the data processing method according to any one of claims 1-8.

Images