DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
The instant application having application No. 18/891,606 filed on September 20, 2024, presents claims 1-20 for examination.
Examiner Notes
Examiner cites particular columns, paragraphs, figures and line numbers in the references as applied to the claims below for the convenience of the applicant. Although the specified citations are representative of the teachings in the art and are applied to the specific limitations within the individual claim, other passages and figures may apply as well. It is respectfully requested that, in preparing responses, the applicant fully consider the references in their entirety as potentially teaching all or part of the claimed invention, as well as the context of the passage as taught by the prior art or disclosed by the examiner.
Claim Objections
Claims 11 and 19-20 are objected to because of the following informalities:
With respect to claim 11, on line 2, “3GPP” should be spelled out in full.
With respect to claim 19, on line 8, “receiving, from the server, code block …” appears to be a typographical error that should recite -- receiving, from the server, a code block --. Claim 20 inherits this deficiency.
Appropriate correction is required.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 2, 3, 4, 5, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Kozhisseri et al. (US 11526334 B2, hereinafter Kozhisseri) in view of Kan et al. “Mobile-LLaMA: Instruction Fine-Tuning Open-Source LLM for Network Analysis in 5G Networks” (hereinafter Kan).
With respect to claim 1, Kozhisseri discloses A method comprising:
scanning, by a processing system including a processor, a plurality of repositories to collect code data, wherein the plurality of repositories includes domain-related code, wherein the plurality of repositories includes public and non-public repositories (e.g., Figs. 1-2 and 4-5 along with associated text, e.g., col. 4:62-col. 5:16, the predetermined code repositories 107 may include one or more external as well as internal databases that store one or more source codes relating to a plurality of functions in a plurality of technologies or programming languages … the source code generator 105 may connect with and retrieve the one or more source codes from the predetermined code repositories 107 using a preconfigured wired and/or wireless communication channel; col. 5:8-12, the predetermined code repositories 107 may include, without limiting to, a public source code repository, a private source code repository, a public issue repository, a private issue repository, online information portals/websites or learning systems; see also claim 1.);
preprocessing, by the processing system, the code data resulting in a structured dataset (e.g., Figs. 1-5 along with associated text, col. 15:3-12, the at least one pre-trained code generation model 106 may be trained by classifying the one or more reference source codes, retrieved from the predetermined repositories, as source code functions and text labels after pre-processing the one or more reference source codes. Further to classification, the source code functions may be converted into an abstract syntax tree structure. Thereafter, a multi-dimensional matrix vector corresponding to the source code functions may be generated based on the abstract syntax tree structure; see also claim 1.);
generating, by the processing system, a Natural Language Processing (NLP) model based on the structured dataset (Id.; col. 12-18, Finally, the at least one pre-trained code generation model 106 may be trained based on the multi-dimensional matrix vector. In an embodiment, the at least one pre-trained code generation model 106 may include, without limitation, a technology learning model, an issue resolution model and a code creation model.);
receiving, by the processing system, input including a natural language statement from equipment of a programmer (e.g., Figs. 1-5 and associated text, e.g., col. 5:22-23, the user inputs 103, which are received in a plain spoken language, such as English; col. 4:38-56, The user 101 may be a developer … the source code generator 105 may receive the user inputs 103 over a preconfigured wired and/or wireless communication channel connecting the source code generator 105 with a user device associated with the user 101; col. 15:45-47, A user may include an application developer, a programmer.), the natural language statement indicating an intent of the programmer to generate code (e.g., Figs. 1-5 and associated text, e.g., col. 6:53-59, The user inputs 103 may include user-specific requirements of the application. An exemplary user input 103 may be that the user 101 is asking the source code generator 105 to “Generate an application to start web logic server and start Oracle database”. As seen from the above example, the objective and/or the requirements of the application are stated in the user inputs 103.);
generating, by the processing system and according to the intent of the programmer, a code block by applying the NLP model to the natural language statement (e.g., Figs. 1-5 and associated text, e.g., col. 13:63-col. 14:1, At block 405, the method 400 includes generating, by the source code generator 105, one or more source codes 213 for the application flow 211 using at least one pre-trained code generation model 106 [NLP model] … the at least one pre-trained code generation model 106 may generate the one or more source codes 213 based on the user inputs 103 … the one or more source codes 213 … may be selected as the one or more best-fit source codes for the application … generating the executable source code 109 may comprise compiling each of the one or more best-fit source codes … and then selecting the one or more best-fit source codes as the executable source code 109 upon successful compilation; see also 5:28-col. 6:15.); and
providing, by the processing system, the code block for presentation at the equipment of the programmer (e.g., Figs. 1-5 and associated text, e.g., col. 12:66-col. 13:2, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109.).
Kozhisseri does not appear to disclose the following, which is taught in analogous art. Kan: network (e.g., Figure 1 on p. 77 and associated text, e.g., p. 79, § Implementation of Mobile-LLaMA, To implement Mobile-LLaMA for 5G NWDAF, we leverage fine-tuning capability of an open-source LLM, LLaMA 2 13B. We need mobile network training data and instructions to enable the 5G network analysis functions … 1) Training Data: For instruction fine-tuning, we collect publicly available 5G network data sets, which align with the operational functions of Mobile-LLaMA within NWDAF … For packet analysis function, we utilize a dataset provided by Deutsche Telekom’s GitHub repository … Utilizing real-world datasets, we have developed a collection of high-quality Python code scripts specifically for 5G data analysis instructions. The primary goal is to use these scripts as a training foundation for the pre-trained Language Model-LLaMA. Each script is crafted to effectively demonstrate the use of essential data analysis libraries, including PyBGPStream for IP routing analysis, Scapy for packet analysis, Pandas for data manipulation, Matplotlib for visualization.) … network domain-specific (e.g., Figs. 1-2 on pp. 78-79, particularly, Figure 2 on p. 79, particularly the prompts depicted in portion (b) “Given a n3.pcap file containing network traffic data, create a Python script using Scapy to process the PCAP …” and portion (c) “Generate Python code to calculate 5G network performance …” [generate network domain-specific code], along with associated text, e.g., p. 78, § Functions of Mobile-LLaMA … The capabilities of Mobile-LLaMA across its three main functions are demonstrated in Fig. 2, and complete prompts, code, and code outputs for each demonstration is available in our GitHub repository).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the invention of Kozhisseri with the invention of Kan because “network management and operation could be performed more efficiently and automatically,” as suggested by Kan (see p. 77, top para.).
With respect to claim 19, Kozhisseri discloses A device (e.g., Fig. 5.), comprising:
a processing system including a processor; and a memory that stores executable instructions that, when executed by the processing system, facilitate performance of operations (e.g., Fig. 5 and associated text, e.g., col. 15:36-45, FIG. 5 illustrates a block diagram of an exemplary computer system 500 for implementing embodiments consistent with the present disclosure. In an embodiment, the computer system 500 may be the source code generator 105 illustrated in FIG. 1 … processor 502 may comprise at least one data processor for executing program components for executing user- or system-generated business processes … processor 502 may be disposed in communication with a memory 505.), the operations comprising:
providing input to a server, the input including a natural language statement (e.g., Figs. 1-2 and 4-5, along with associated text, e.g., col. 4, 57-59, the source code generator 105 may be a computing system such as … a server; col. 5:22-23, the user inputs 103, which are received in a plain spoken language, such as English … the source code generator 105 may generate the one or more source codes for the application flow using at least one of a pre-trained code generation model 106 configured in the source code generator 105. For example, the at least one pre-trained code generation model 106 may be the neural network models that are trained to comprehend the user inputs 103.), the natural language statement indicating an intent by a programmer to generate code (e.g., Figs. 1, 2, and 4, along with associated text, e.g., col. 6:53-59, The user inputs 103 may include user-specific requirements of the application. An exemplary user input 103 may be that the user 101 is asking the source code generator 105 to “Generate an application to start web logic server and start Oracle database”. As seen from the above example, the objective and/or the requirements of the application are stated in the user inputs 103.);
receiving, from the server, code block, wherein the code block is generated based on applying an NLP model to the natural language statement (e.g., Figs. 1-5 and associated text, e.g., col. 13:63-col. 14:1, At block 405, the method 400 includes generating, by the source code generator 105, one or more source codes 213 for the application flow 211 using at least one pre-trained code generation model 106 [NLP model] … the at least one pre-trained code generation model 106 may generate the one or more source codes 213 based on the user inputs 103 … the one or more source codes 213 … may be selected as the one or more best-fit source codes for the application … generating the executable source code 109 may comprise compiling each of the one or more best-fit source codes … and then selecting the one or more best-fit source codes as the executable source code 109 upon successful compilation; see also 5:28-col. 6:15.), wherein the NLP model is generated from training based on code data (e.g., Figs. 1-2 and 4-5, along with associated text, e.g., col. 15:2-15, the at least one pre-trained code generation model 106 may be trained by classifying the one or more reference source codes, retrieved from the predetermined repositories, as source code functions and text labels after pre-processing the one or more reference source codes. Further to classification, the source code functions may be converted into an abstract syntax tree structure. Thereafter, a multi-dimensional matrix vector corresponding to the source code functions may be generated based on the abstract syntax tree structure. Finally, the at least one pre-trained code generation model 106 may be trained based on the multi-dimensional matrix vector; col. 7:15-20, the at least one pre-trained code generation model 106 configured … based on …one or more reference source codes retrieved from the predetermined code repositories 107. ; see also claim 1.), wherein the code data is scanned from a plurality of repositories that include domain-related code (e.g., Figs. 1-2 and 4-5 along with associated text, e.g., col. 4:62-col. 5:16, the predetermined code repositories 107 may include one or more external as well as internal databases that store one or more source codes relating to a plurality of functions in a plurality of technologies or programming languages … the source code generator 105 may connect with and retrieve the one or more source codes from the predetermined code repositories 107 using a preconfigured wired and/or wireless communication channel.); col. 15:3-8, the at least one pre-trained code generation model 106 may be trained by classifying the one or more reference source codes, retrieved from the predetermined repositories, as source code functions and text labels after pre-processing the one or more reference source codes; see also claim 1.); and
providing, to the server, at least one of feedback data, additional code data, or a combination thereof (e.g., Figs. 1-5 and associated text, e.g., col. 12:66-13:2, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109; col. 15:30-24, As indicated in block 411 of the method 400, the code generation model 106 of the source code generator 105 may undergo a reinforcement learning and/or relearning based on the user feedback on the executable source code 109; col. 13:3-23, In case, the user 101 has “fully accepted” the given executable source code 109 [additional code data], the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106.), wherein the feedback data is associated with an analysis of the code block (Id., particularly, col. 12:66-13:2, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109; col. 12:66-col. 13:5, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109. As an example, the user feedback may be at least one that the user 101 has “fully accepted”, “partially accepted” or “not accepted” the given executable source code 109.), wherein the additional code data includes the code block (e.g., Figs. 1-5 and associated text, e.g., col. 13:3-23, In case, the user 101 has “fully accepted” the given executable source code 109 [additional code data], the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106.), and wherein the NLP model is adjusted based on at least one of the feedback data, the additional code data, or a combination thereof resulting in an adjusted NLP model (e.g., Figs. 1-5 and associated text, e.g., col. 15:30-24, As indicated in block 411 of the method 400, the code generation model 106 of the source code generator 105 may undergo a reinforcement learning and/or relearning based on the user feedback on the executable source code 109; col. 13:3-23, In case, the user 101 has “fully accepted” the given executable source code 109 [additional code data], the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106; col. 13:3-23, In case, the user 101 has “fully accepted” the given executable source code 109, the selected executable source code 109 [additional code data] may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106. Also, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements. On the other hand, if the given executable source code 109 [additional code data] is “partially accepted” by the user 101, then the partially accepted part of the source code may be passed to the pre-trained code generation model 106 for the reinforcement learning and to learn the mistake. Also, a penalty may be assigned to the pre-trained code generation model 106, so that the pre-trained code generation model 106 ensures not to retrieve same or similar executable source codes 109 for the similar application requirements.).
Kozhisseri does not appear to disclose the following, which is taught in analogous art. Kan: network domain-specific (e.g., Figs. 1-2 on pp. 78-79, particularly, Figure 2 on p. 79, particularly the prompts depicted in portion (b) “Given a n3.pcap file containing network traffic data, create a Python script using Scapy to process the PCAP …” and portion (c) “Generate Python code to calculate 5G network performance …” [generate network domain-specific code], along with associated text, e.g., p. 78, § Functions of Mobile-LLaMA … The capabilities of Mobile-LLaMA across its three main functions are demonstrated in Fig. 2, and complete prompts, code, and code outputs for each demonstration is available in our GitHub repository) … network (e.g., Figure 1 on p. 77 and associated text, e.g., p. 79, § Implementation of Mobile-LLaMA, To implement Mobile-LLaMA for 5G NWDAF, we leverage fine-tuning capability of an open-source LLM, LLaMA 2 13B. We need mobile network training data and instructions to enable the 5G network analysis functions … 1) Training Data: For instruction fine-tuning, we collect publicly available 5G network data sets, which align with the operational functions of Mobile-LLaMA within NWDAF … For packet analysis function, we utilize a dataset provided by Deutsche Telekom’s GitHub repository … Utilizing real-world datasets, we have developed a collection of high-quality Python code scripts specifically for 5G data analysis instructions. The primary goal is to use these scripts as a training foundation for the pre-trained Language Model-LLaMA. Each script is crafted to effectively demonstrate the use of essential data analysis libraries, including PyBGPStream for IP routing analysis, Scapy for packet analysis, Pandas for data manipulation, Matplotlib for visualization.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the invention of Kozhisseri with the invention of Kan because “network management and operation could be performed more efficiently and automatically,” as suggested by Kan (see p. 77, top para.).
With respect to claim 2, Kozhisseri also discloses receiving, by the processing system, feedback data from the equipment of the programmer, the feedback data being associated with an analysis of the code block by the programmer (e.g., Figs. 1-5 and associated text, e.g., col. 12:66-13:2, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109; col. 15:30-24, As indicated in block 411 of the method 400, the code generation model 106 of the source code generator 105 may undergo a reinforcement learning and/or relearning based on the user feedback on the executable source code 109.); and
adjusting, by the processing system, the NLP model based on the feedback data resulting in an adjusted NLP model (Id.; col. 13:3-23, the user feedback may be at least one that the user 101 has “fully accepted”, “partially accepted” or “not accepted” the given executable source code 109. In case, the user 101 has “fully accepted” the given executable source code 109, the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106. Also, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements. On the other hand, if the given executable source code 109 is “partially accepted” by the user 101, then the partially accepted part of the source code may be passed to the pre-trained code generation model 106 for the reinforcement learning and to learn the mistake. Also, a penalty may be assigned to the pre-trained code generation model 106, so that the pre-trained code generation model 106 ensures not to retrieve same or similar executable source codes 109 for the similar application requirements. Similar learning strategy is followed when the user feedback is “not accepted”.).
With respect to claim 3, Kozhisseri also discloses receiving, by the processing system, a second input including a second natural language statement from the equipment of the programmer (e.g., Figs. 1-5 and associated text, e.g., col. 5:22-23, the user inputs 103, which are received in a plain spoken language, such as English; col. 4:38-56, The user 101 may be a developer … the source code generator 105 may receive the user inputs 103 over a preconfigured wired and/or wireless communication channel connecting the source code generator 105 with a user device associated with the user 101; col. 15:45-47, A user may include an application developer, a programmer; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 13:10-13, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements [code generation is repeated and reinforcement learning adjusts model behavior for subsequent code generation, i.e., second]; see also col. 5:58-62, col. 14:45-50, and col. 15:30-33.), the natural language statement indicating a second intent of the programmer to generate second code (e.g., Figs. 1-5 and associated text, e.g., col. 6:53-59, The user inputs 103 may include user-specific requirements of the application. An exemplary user input 103 may be that the user 101 is asking the source code generator 105 to “Generate an application to start web logic server and start Oracle database”. As seen from the above example, the objective and/or the requirements of the application are stated in the user inputs 103; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 13:10-13, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements [code generation is repeated and reinforcement learning adjusts model behavior for subsequent code generation, i.e., second]; see also col. 5:58-62, col. 14:45-50, and col. 15:30-33.); generating, by the processing system and according to the second intent of the programmer, a second code block by applying the adjusted NLP model to the second natural language statement (e.g., Figs. 1-5 and associated text, e.g., col. 13:63-col. 14:1, At block 405, the method 400 includes generating, by the source code generator 105, one or more source codes 213 for the application flow 211 using at least one pre-trained code generation model 106 [NLP model] … the at least one pre-trained code generation model 106 may generate the one or more source codes 213 based on the user inputs 103 … the one or more source codes 213 … may be selected as the one or more best-fit source codes for the application … generating the executable source code 109 may comprise compiling each of the one or more best-fit source codes … and then selecting the one or more best-fit source codes as the executable source code 109 upon successful compilation; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 13:10-13, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements [code generation is repeated and reinforcement learning adjusts model behavior for subsequent code generation, i.e., generating, by the processing system and according to the second intent of the programmer, a second code block by applying the adjusted NLP model to the second natural language statement]; see also col. 5:58-62, col. 14:45-50, and col. 15:30-33.); and providing, by the processing system, the second code block for presentation at the equipment of the programmer (e.g., Figs. 1-5 and associated text, e.g., col. 12:66-col. 13:2, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109; col. 12:12-14, the historical information related to the one or more executable codes, obtained from user feedback; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 12:12-14, the historical information related to the one or more executable codes, obtained from user feedback [the feedback process is repeated, i.e., providing, by the processing system, the second code block for presentation at the equipment of the programmer]; see also col. 5:58-62, col. 14:45-50, and col. 15:30-33.) and Kan further teaches network domain-specific (e.g., Figs. 1-2 on pp. 78-79, particularly, Figure 2 on p. 79, particularly the prompts depicted in portion (b) “Given a n3.pcap file containing network traffic data, create a Python script using Scapy to process the PCAP …” and portion (c) “Generate Python code to calculate 5G network performance …” [generate network domain-specific code], along with associated text, e.g., p. 78, § Functions of Mobile-LLaMA … The capabilities of Mobile-LLaMA across its three main functions are demonstrated in Fig. 2, and complete prompts, code, and code outputs for each demonstration is available in our GitHub repository).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the invention of Kozhisseri with the invention of Kan for the same reason set forth above.
With respect to claim 4, Kozhisseri also discloses receiving, by the processing system, additional code data that includes the code block (e.g., Figs. 1-4 and associated text, e.g., col. 13:3-23, In case, the user 101 has “fully accepted” the given executable source code 109 [additional code data that includes the code block], the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106.; col. 15:30-24, As indicated in block 411 of the method 400, the code generation model 106 of the source code generator 105 may undergo a reinforcement learning and/or relearning based on the user feedback on the executable source code 109.);
scanning, by the processing system, the additional code data (Id., particularly, the selected executable source code 109 [additional code block] may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106.); and
generating, by the processing system, an additional NLP model based on at least a portion of the structured dataset and at least a portion of the additional code data (Id.; col. 15:3-8, the at least one pre-trained code generation model 106 [NLP model] may be trained by classifying the one or more reference source codes, retrieved from the predetermined repositories [the code data]; col. 13:3-23, In case, the user 101 has “fully accepted” the given executable source code 109, the selected executable source code 109 [additional code data] may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106. Also, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements. On the other hand, if the given executable source code 109 [additional code data] is “partially accepted” by the user 101, then the partially accepted part of the source code may be passed to the pre-trained code generation model 106 for the reinforcement learning and to learn the mistake. Also, a penalty may be assigned to the pre-trained code generation model 106, so that the pre-trained code generation model 106 ensures not to retrieve same or similar executable source codes 109 for the similar application requirements [code generation model 106 is retrained using the generated executable source code 109, i.e., additional code data, thereby creating a changed code generation model, i.e., generating, by the processing system, an additional NLP model based on at least a portion of the additional code data, and the executable source code 109 used for the retraining was generated by the code generation model 106, which was pre-trained on the code data retrieved from the repositories, i.e., generating an additional NLP model based on at least a portion of the code data]; see also col. 15:30-24.).
With respect to claim 5, Kozhisseri also discloses receiving, by the processing system, a second input including a second natural language statement from the equipment of the programmer (e.g., Figs. 1-5 and associated text, e.g., col. 5:22-23, the user inputs 103, which are received in a plain spoken language, such as English; col. 4:38-56, The user 101 may be a developer … the source code generator 105 may receive the user inputs 103 over a preconfigured wired and/or wireless communication channel connecting the source code generator 105 with a user device associated with the user 101; col. 15:45-47, A user may include an application developer, a programmer; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 13:10-13, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements [code generation is repeated and reinforcement learning adjusts model behavior for subsequent code generation, i.e., second]; see also col. 5:58-62, col. 14:45-50, and col. 15:30-33.), the natural language statement indicating a second intent of the programmer to generate second code (e.g., Figs. 1-5 and associated text, e.g., col. 6:53-59, The user inputs 103 may include user-specific requirements of the application. An exemplary user input 103 may be that the user 101 is asking the source code generator 105 to “Generate an application to start web logic server and start Oracle database”. As seen from the above example, the objective and/or the requirements of the application are stated in the user inputs 103; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 13:10-13, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements [code generation is repeated and reinforcement learning adjusts model behavior for subsequent code generation, i.e., second]; see also col. 5:58-62, col. 14:45-50, and col. 15:30-33.); generating, by the processing system and according to the second intent of the programmer, a second code block by applying the additional NLP model to the second natural language statement (e.g., Figs. 1-5 and associated text, e.g., col. 13:63-col. 14:1, At block 405, the method 400 includes generating, by the source code generator 105, one or more source codes 213 for the application flow 211 using at least one pre-trained code generation model 106 [NLP model] … the at least one pre-trained code generation model 106 may generate the one or more source codes 213 based on the user inputs 103 … the one or more source codes 213 … may be selected as the one or more best-fit source codes for the application … generating the executable source code 109 may comprise compiling each of the one or more best-fit source codes … and then selecting the one or more best-fit source codes as the executable source code 109 upon successful compilation; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 13:10-13, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements [code generation is repeated and reinforcement learning adjusts model behavior for subsequent code generation, i.e., generating, by the processing system and according to the second intent of the programmer, a second code block by applying the additional NLP model to the second natural language statement]; see also col. 5:58-62, col. 14:45-50, and col. 15:30-33.); and providing, by the processing system, the second code block for presentation at the equipment of the programmer (e.g., Figs. 1-5 and associated text, e.g., col. 12:66-col. 13:2, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109; col. 12:12-14, the historical information related to the one or more executable codes, obtained from user feedback; col. 12:35-44, after obtaining each of the one or more executable source codes 109 from the validation module 225, the recommendation engine 226 in the source code generator 105 may check if any previous history exists with similar application; col. 12:12-14, the historical information related to the one or more executable codes, obtained from user feedback [the feedback process is repeated, i.e., providing, by the processing system, the second code block for presentation at the equipment of the programmer]; see also col. 5:58-62, col. 14:45-50, and col. 15:30-33.) and Kan further teaches network domain-specific (e.g., Figs. 1-2 on pp. 78-79, particularly, Figure 2 on p. 79, particularly the prompts depicted in portion (b) “Given a n3.pcap file containing network traffic data, create a Python script using Scapy to process the PCAP …” and portion (c) “Generate Python code to calculate 5G network performance …” [generate network domain-specific code], along with associated text, e.g., p. 78, § Functions of Mobile-LLaMA … The capabilities of Mobile-LLaMA across its three main functions are demonstrated in Fig. 2, and complete prompts, code, and code outputs for each demonstration is available in our GitHub repository).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the invention of Kozhisseri with the invention of Kan for the same reason set forth above.
Claims 6 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Kozhisseri in view of Kan, as applied to claims 5 and 19 above, and further in view of Dohmke, “GitHub Copilot Enterprise is now generally available” (hereinafter Dohmke).
With respect to claim 6, Kozhisseri also discloses wherein use of the additional NLP model is
Kozhisseri does not appear to disclose the following, which is taught in analogous art, Dohmke: limited to an entity associated with the programmer or any entity authorized by the programmer (e.g., p. 9, top para., Built with the world's leading large language model, customized to your organization, and deeply integrated into GitHub's surfaces, GitHub Copilot Enterprise brings immense value to every organization; p. 9, 4th -6th paras., GitHub Copilot Enterprise comes with the same seat and policy management features as Copilot Business … If you're an enterprise administrator: you can manage access for organizations within your enterprise, while organization administrators can handle access for teams and individuals within their organization. If you're a developer: once you're assigned a GitHub Copilot Enterprise seat, you'll automatically see Copilot in GitHub Enterprise.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Dohmke because NLP models can have access to private data and the invention of Dohmke would help ensure that such private remains private by using an access management policy.
With respect to claim 20, Kozhisseri also discloses wherein use of the adjusted NLP model is (e.g., Figs. 1-5 and associated text, e.g., col. 4:38-43, The user 101 may be a developer; col. 15:45-47, A user may include an application developer, a programmer; col. 15:30-24, As indicated in block 411 of the method 400, the code generation model 106 of the source code generator 105 may undergo a reinforcement learning and/or relearning based on the user feedback on the executable source code 109.).
Kozhisseri does not appear to disclose the following, which is taught in analogous art, Dohmke: limited to an entity associated with the programmer or any entity authorized by the programmer (e.g., p. 9, top para., Built with the world's leading large language model, customized to your organization, and deeply integrated into GitHub's surfaces, GitHub Copilot Enterprise brings immense value to every organization; p. 9, 4th -6th paras., GitHub Copilot Enterprise comes with the same seat and policy management features as Copilot Business … If you're an enterprise administrator: you can manage access for organizations within your enterprise, while organization administrators can handle access for teams and individuals within their organization. If you're a developer: once you're assigned a GitHub Copilot Enterprise seat, you'll automatically see Copilot in GitHub Enterprise.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Dohmke because NLP models can have access to private data and the invention of Dohmke would help ensure that such private remains private by using an access management policy.
Claims 7-10 are rejected under 35 U.S.C. 103 as being unpatentable over Kozhisseri in view of Kan, as applied to claim 1 above, and further in view of Hoban et al. (US 20250321719 A1, hereinafter Hoban).
With respect to claim 7, Kozhisseri does not appear to disclose the following, which is taught in analogous art, Hoban: wherein the generating the code block comprises generating multiple versions of the code block (e.g., Figs. 1-4 and associated text, e.g., [0004], The system receives a natural language request for generating IaC for configuring a computing infrastructure on a target cloud platform. The system generates a prompt requesting a machine learning based language model to generate IaC using a configuration language; [0006] the system may build the same computing infrastructure specified using natural language on different cloud platforms or a combination of multiple cloud platforms; [0027], a solution for a user to generate IaC using natural language … The IaC generated may use different configuration languages; claim 1, receiving … a natural language request for generating infrastructure-as-code for configuring a computing infrastructure using a target cloud platform … receiving, from the machine learning based language model, infrastructure-as-code specified using the configuration language; claim 8, wherein the configuration language is a first configuration language … wherein the natural language request is a particular natural language request … receiving … the particular natural language request for generating infrastructure-as-code for configuring a computing infrastructure using the target cloud platform using a second configuration language; see also claim 7.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Hoban because it “provide[s] several technical advantages. For example, the systems and methods disclosed herein provide a solution for a user to generate IaC using natural platforms, managing and provisioning resources on various cloud platforms” and it “provides an improved user experience and simplifies the user interface available to users who want to configure infrastructure on cloud platforms”, as suggested by Hoban (see [0027] and [0029]).
With respect to claim 8, Hoban further teaches wherein the multiple versions of the code block are applicable to different vendor equipment (e.g., Figs. 1-4 and associated text, e.g., [0004], The system receives a natural language request for generating IaC for configuring a computing infrastructure on a target cloud platform. The system generates a prompt requesting a machine learning based language model to generate IaC using a configuration language; [0006] the system may build the same computing infrastructure specified using natural language on different cloud platforms or a combination of multiple cloud platforms; [0027], a solution for a user to generate IaC using natural language … The IaC generated may use different configuration languages; claim 1, receiving … a natural language request for generating infrastructure-as-code for configuring a computing infrastructure using a target cloud platform … receiving, from the machine learning based language model, infrastructure-as-code specified using the configuration language; claim 7, wherein the target cloud platform is a first target cloud platform, wherein the prompt is a first prompt, wherein the natural language request is a particular natural language request, the computer-implemented method further comprising: receiving, via the user interface, the particular natural language request for generating infrastructure-as-code for configuring a computing infrastructure using a second target cloud platform using a configuration language.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Hoban for the same reason set forth above.
With respect to claim 9, Hoban further teaches wherein the multiple versions of the code block are in different programming languages (e.g., Figs. 1-4 and associated text, e.g., [0004], The system receives a natural language request for generating IaC for configuring a computing infrastructure on a target cloud platform. The system generates a prompt requesting a machine learning based language model to generate IaC using a configuration language; [0006] the system may build the same computing infrastructure specified using natural language on different cloud platforms or a combination of multiple cloud platforms; [0027], a solution for a user to generate IaC using natural language … The IaC generated may use different configuration languages; claim 1, receiving … a natural language request for generating infrastructure-as-code for configuring a computing infrastructure using a target cloud platform … receiving, from the machine learning based language model, infrastructure-as-code specified using the configuration language; claim 8, wherein the configuration language is a first configuration language … wherein the natural language request is a particular natural language request … receiving … the particular natural language request for generating infrastructure-as-code for configuring a computing infrastructure using the target cloud platform using a second configuration language.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Hoban for the same reason set forth above.
With respect to claim 10, Hoban further teaches wherein the multiple versions of the code block are applicable to different communications service providers (e.g., Figs. 1-4 and associated text, e.g., [0036], Cloud platform 135 may provide various cloud computing services … The cloud platform 135 may be referred to herein as cloud resource provider … the techniques disclosed herein are applicable to any type of resource providers, including physical resource providers … Physical resources are resources offered and managed by physical resource providers, including but are not limited to … networking; [0004], The system receives a natural language request for generating IaC for configuring a computing infrastructure on a target cloud platform. The system generates a prompt requesting a machine learning based language model to generate IaC using a configuration language; [0006] the system may build the same computing infrastructure specified using natural language on different cloud platforms or a combination of multiple cloud platforms; see also [0002], [0031], claim 1, and claim 7.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Hoban for the same reason set forth above.
Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Kozhisseri in view of Kan, as applied to claim 1 above, and further in view of Zou et al. “TelecomGPT: A Framework to Build Telecom-Specfic Large Language Models” (hereinafter Zou).
With respect to claim 11, Kozhisseri also discloses obtaining, by the processing system, information ; and adjusting, by the processing system, the NLP model based on the information resulting in an adjusted NLP model (Id.; col. 13:3-23, the user feedback may be at least one that the user 101 has “fully accepted”, “partially accepted” or “not accepted” the given executable source code 109. In case, the user 101 has “fully accepted” the given executable source code 109, the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106. Also, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements. On the other hand, if the given executable source code 109 is “partially accepted” by the user 101, then the partially accepted part of the source code may be passed to the pre-trained code generation model 106 for the reinforcement learning and to learn the mistake. Also, a penalty may be assigned to the pre-trained code generation model 106, so that the pre-trained code generation model 106 ensures not to retrieve same or similar executable source codes 109 for the similar application requirements. Similar learning strategy is followed when the user feedback is “not accepted”.).
Kozhisseri does not appear to disclose the following, which is taught in analogous art, Zou: indicating a revision to a 3GPP standard (e.g., Fig. 1 on p. 5 and associated text, e.g., p. 2, right col., Frequent knowledge updating: regular release announcement from SDO … make it difficult for LLMs, with their high training cost and long training time, to be updated in a timely manner. Therefore, instead of pre-training a Telecom LLM from scratch which is expensive and unprepared for the time being, it would be efficient and reasonable to consider adapting general purpose LLMs to telecom domain under acceptable cost and training time, which is exactly the target of this paper; p. 4, § Dataset, Following our training method, we need to build three datasets, namely the pre-training dataset for continual pretraining … The pre-training dataset is collected mainly from web, which includes Telecom standards … 3GPP is the main SDO in the area of Telecommunication … We scrap [sic] the technical specifications, reports, and documents from release 8 to 19 on 3GPP FTP site.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Zou because “it would be efficient and reasonable to consider adapting general purpose LLMs to telecom domain under acceptable cost and training time”, as suggested by Zou (see p. 2, right col., 1st full para.).
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Kozhisseri in view of Kan, as applied to claim 1 above, and further in view of Padmanabhan (US 20220199075 A1, hereinafter Padmanabhan).
With respect to claim 12, Kozhisseri also discloses obtaining, by the processing system, information indicating ; and adjusting, by the processing system, the NLP model based on the information resulting in an adjusted NLP model (Id.; col. 13:3-23, the user feedback may be at least one that the user 101 has “fully accepted”, “partially accepted” or “not accepted” the given executable source code 109. In case, the user 101 has “fully accepted” the given executable source code 109, the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106. Also, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements. On the other hand, if the given executable source code 109 is “partially accepted” by the user 101, then the partially accepted part of the source code may be passed to the pre-trained code generation model 106 for the reinforcement learning and to learn the mistake. Also, a penalty may be assigned to the pre-trained code generation model 106, so that the pre-trained code generation model 106 ensures not to retrieve same or similar executable source codes 109 for the similar application requirements. Similar learning strategy is followed when the user feedback is “not accepted”.).
Kozhisseri does not appear to disclose the following, which is taught in analogous art, Padmanabhan: a change to a vendor equipment (e.g., [0002], SDNs with multi-vendor equipment; [0054], A SDN controller vendor will obtain and use a training set to train the language processing engine 222 to achieve the ability to convert text inputs into specific control actions and commands … Next, the language processing engine 222 may be trained to process API documents in human language. This may include collecting all API documents associated with the SDN controller, and using these documents to learn various features, capabilities, and parameters associated with the API's … Further, the language processing engine 222 may use adaptive learning to be updated with new API's as they become available or are used in the SDN.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Padmanabhan because “Using machine learning models and intent based networking, a user may more naturally manage and control a SDN, providing many of the benefits described above and avoiding many of the challenges also described above”, as suggested Padmanabhan (see [0110]).
Claims 13-14 are rejected under 35 U.S.C. 103 as being unpatentable over Kozhisseri in view of Kan, as applied to claim 1 above, and further in view of Arumugam Selvaraj (US 20230418565 A1, hereinafter Selvaraj).
With respect to claim 13, Kozhisseri does not appear to disclose the following, which is taught in analogous art, Selvaraj: providing, by the processing system to the equipment of the programmer, suggestions for code improvements based on the code block (e.g., Figs. 2-3 and 5 and associated text, e.g., [0034], Code suggestion 213 may use generative models, machine learning models such as Generative Pre-trained Transformer (GPT), trained to generate code suggestions; [0061], The code suggestion feature 213 of code development service 210 may analyze the entered characters to determine a code suggestion 520, which may be displayed and added, as indicated at 522; see also [0031].).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Selvaraj because it “may improve the user experience”, as suggested by Selvaraj (see [0020]).
With respect to claim 14, Kozhisseri does not appear to disclose the following, which is taught in analogous art, Selvaraj: wherein the providing the code block comprises integrating, by the processing system, the code block into an Integrated Development Environment (IDE) used by the programmer (e.g., Figs. 2-3 and 5 and associated text, e.g., [0061], Integrated development environment interface 500 may be implemented on a client of the code development service 210, as depicted in FIG. 2, or hosted as part of the code development service 210, as depicted in FIG. 2. Integrated development environment interface 500 may implement a code editor 510 (e.g., a text editor) which may allow a user to enter code in a programming language. The code suggestion feature 213 of code development service 210 may analyze the entered characters to determine a code suggestion 520, which may be displayed and added, as indicated at 522.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Selvaraj because it “may improve the user experience”, as suggested by Selvaraj (see [0020]).
Claims 15-18 are rejected under 35 U.S.C. 103 as being unpatentable over Kozhisseri et al. (US 11526334 B2, hereinafter Kozhisseri) in view of Kan et al. “Mobile-LLaMA: Instruction Fine-Tuning Open-Source LLM for Network Analysis in 5G Networks” (hereinafter Kan), Vadaparty et al. (US 12360791 B1, hereinafter Vadaparty), Hoban et al. (US 20250321719 A1, hereinafter Hoban).
With respect to claim 15, Kozhisseri discloses A non-transitory machine-readable medium, comprising executable instructions that, when executed by a processing system including a processor, facilitate performance of operations (e.g., Fig. 5 and associated text, e.g., col. 2:36-41, a non-transitory computer readable medium including instructions stored thereon that when processed by at least one processor cause a source code generator to perform operations comprising dynamically generating an executable source code for an application.), the operations comprising:
obtaining code data from a plurality of repositories, wherein each of the plurality of repositories includes domain-related code (e.g., Figs. 1-2 and 4-5 along with associated text, e.g., col. 4:62-col. 5:16, the predetermined code repositories 107 may include one or more external as well as internal databases that store one or more source codes relating to a plurality of functions in a plurality of technologies or programming languages … the source code generator 105 may connect with and retrieve the one or more source codes from the predetermined code repositories 107 using a preconfigured wired and/or wireless communication channel; see also claim 1.);
generating a Natural Language Processing (NLP) model based on the code data (e.g., Figs. 1-2 and 4-5, along with associated text, e.g., col. 15:2-15, the at least one pre-trained code generation model 106 may be trained by classifying the one or more reference source codes, retrieved from the predetermined repositories, as source code functions and text labels after pre-processing the one or more reference source codes. Further to classification, the source code functions may be converted into an abstract syntax tree structure. Thereafter, a multi-dimensional matrix vector corresponding to the source code functions may be generated based on the abstract syntax tree structure. Finally, the at least one pre-trained code generation model 106 may be trained based on the multi-dimensional matrix vector; see also claim 1.);
receiving input , the input including a natural language statement from equipment of a programmer (Id.; col. 5:22-23, the user inputs 103, which are received in a plain spoken language, such as English; col. 4:38-56, The user 101 may be a developer … the source code generator 105 may receive the user inputs 103 over a preconfigured wired and/or wireless communication channel connecting the source code generator 105 with a user device associated with the user 101; col. 15:45-47, A user may include an application developer, a programmer.), the natural language statement indicating an intent of the programmer to generate code (e.g., Figs. 1, 2, and 4, along with associated text, e.g., col. 6:53-59, The user inputs 103 may include user-specific requirements of the application. An exemplary user input 103 may be that the user 101 is asking the source code generator 105 to “Generate an application to start web logic server and start Oracle database”. As seen from the above example, the objective and/or the requirements of the application are stated in the user inputs 103.);
generating according to the intent of the programmer, a code block by applying the NLP model to the natural language statement (e.g., Figs. 1-5 and associated text, e.g., col. 13:63-col. 14:1, At block 405, the method 400 includes generating, by the source code generator 105, one or more source codes 213 for the application flow 211 using at least one pre-trained code generation model 106 [NLP model] … the at least one pre-trained code generation model 106 may generate the one or more source codes 213 based on the user inputs 103 … the one or more source codes 213 … may be selected as the one or more best-fit source codes for the application … generating the executable source code 109 may comprise compiling each of the one or more best-fit source codes … and then selecting the one or more best-fit source codes as the executable source code 109 upon successful compilation; see also 5:28-col. 6:15.), ; and
providing the code block for presentation at the equipment of the programmer (e.g., Figs. 1-5 and associated text, e.g., col. 12:66-col. 13:2, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109.).
Kozhisseri does not appear to disclose the following, which is taught in analogous art, Kan: a particular (e.g., Figure 1 on p. 77 and associated text, e.g., p. 79, § Implementation of Mobile-LLaMA, To implement Mobile-LLaMA for 5G NWDAF, we leverage fine-tuning capability of an open-source LLM, LLaMA 2 13B. We need mobile network training data and instructions to enable the 5G network analysis functions … 1) Training Data: For instruction fine-tuning, we collect publicly available 5G network data sets, which align with the operational functions of Mobile-LLaMA within NWDAF … For packet analysis function, we utilize a dataset provided by Deutsche Telekom’s GitHub repository … Utilizing real-world datasets, we have developed a collection of high-quality Python code scripts specifically for 5G data analysis instructions. The primary goal is to use these scripts as a training foundation for the pre-trained Language Model-LLaMA. Each script is crafted to effectively demonstrate the use of essential data analysis libraries, including PyBGPStream for IP routing analysis, Scapy for packet analysis, Pandas for data manipulation, Matplotlib for visualization.) … a particular domain-specific … compatible with the particular domain-related code (Id.; Figure 2 on p. 79, particularly the prompts depicted in portion (b) “Given a n3.pcap file containing network traffic data, create a Python script using Scapy to process the PCAP …” and portion (c) “Generate Python code to calculate 5G network performance …”, along with associated text, e.g., p. 78, § Functions of Mobile-LLaMA … The capabilities of Mobile-LLaMA across its three main functions are demonstrated in Fig. 2, and complete prompts, code, and code outputs for each demonstration is available in our GitHub repository; p. 82, § Discussion, Our instruction fine-tuning has enabled Mobile LLaMA to produce quality code; see also p. 81, § Metric for Network Code Evaluation.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the invention of Kozhisseri with the invention of Kan because “network management and operation could be performed more efficiently and automatically,” as suggested by Kan (see p. 77, top para.).
Kozhisseri does not appear to disclose the following, which is taught in analogous art, Vadaparty: via an Application Programming Interface (API) (e.g., Fig. 1 and associated text, e.g., col. 5:3-6, The API gateway 19 can route the user's inputs/prompts to the LLM 14 and, once the LLM 14 generates a response, route the response to the front end interface 17 for sending on to the user device 16.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Vadaparty because it can result in “simplifying the interaction” between clients and the model, as suggested by Vadaparty (see col. 5:7-8).
Kozhisseri does not appear to disclose the following, which is taught in analogous art, Hoban: wherein the generating the code block comprises generating multiple versions of the code block, and wherein the multiple versions of the code block are at least one of applicable to different vendor equipment, in different programming languages, applicable to different service providers, or a combination thereof … the multiple versions of (e.g., e.g., Figs. 1-4 and associated text, e.g., [0004], The system receives a natural language request for generating IaC for configuring a computing infrastructure on a target cloud platform. The system generates a prompt requesting a machine learning based language model to generate IaC using a configuration language ... The response includes the IaC specified using the configuration language … The system sends the IaC for displaying via a user interface; [0006] the system may build the same computing infrastructure specified using natural language on different cloud platforms or a combination of multiple cloud platforms; [0027], a solution for a user to generate IaC using natural language … The IaC generated may use different configuration languages; claim 1, receiving … a natural language request for generating infrastructure-as-code for configuring a computing infrastructure using a target cloud platform … receiving, from the machine learning based language model, infrastructure-as-code specified using the configuration language; and displaying … the infrastructure-as-code specified using the configuration language; claim 7, wherein the target cloud platform is a first target cloud platform … wherein the natural language request is a particular natural language request … receiving … the particular natural language request for generating infrastructure-as-code for configuring a computing infrastructure using a second target cloud platform; claim 8, wherein the configuration language is a first configuration language … wherein the natural language request is a particular natural language request … receiving … the particular natural language request for generating infrastructure-as-code for configuring a computing infrastructure using the target cloud platform using a second configuration language … displaying … the infrastructure-as-code specified using the second configuration language.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the invention of Kozhisseri with the invention of Hoban because it “provide[s] several technical advantages. For example, the systems and methods disclosed herein provide a solution for a user to generate IaC using natural platforms, managing and provisioning resources on various cloud platforms” and it “provides an improved user experience and simplifies the user interface available to users who want to configure infrastructure on cloud platforms”, as suggested by Hoban (see [0027] and [0029]).
With respect to claim 16, Kozhisseri also discloses wherein the plurality of repositories includes public and non-public repositories, and wherein (e.g., Fig. 1 and associated text, e.g., col. 5:8-12, the predetermined code repositories 107 may include, without limiting to, a public source code repository, a private source code repository, a public issue repository, a private issue repository, online information portals/websites or learning systems.) and Kan further teaches the particular domain-related code is a network domain-related code (e.g., Figure 1 on p. 77 and associated text, e.g., p. 79, § Implementation of Mobile-LLaMA, To implement Mobile-LLaMA for 5G NWDAF, we leverage fine-tuning capability of an open-source LLM, LLaMA 2 13B. We need mobile network training data and instructions to enable the 5G network analysis functions … 1) Training Data: For instruction fine-tuning, we collect publicly available 5G network data sets, which align with the operational functions of Mobile-LLaMA within NWDAF … For packet analysis function, we utilize a dataset provided by Deutsche Telekom’s GitHub repository … Utilizing real-world datasets, we have developed a collection of high-quality Python code scripts specifically for 5G data analysis instructions. The primary goal is to use these scripts as a training foundation for the pre-trained Language Model-LLaMA.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the invention of Kan for the same reason set forth above.
With respect to claim 17, Kozhisseri also discloses also discloses receiving feedback data from the equipment of the programmer, the feedback data being associated with an analysis of the code block by the programmer (e.g., Figs. 1-4 and associated text, e.g., col. 12:66-13:2, after recommending the executable source code 109 to the user 101, the source code generator 105 may collect reviews and feedback from the user 101 on the presented executable source code 109; col. 15:30-24, As indicated in block 411 of the method 400, the code generation model 106 of the source code generator 105 may undergo a reinforcement learning and/or relearning based on the user feedback on the executable source code 109.); and adjusting the NLP model based on the feedback data resulting in an adjusted NLP model (Id.; col. 13:3-23, the user feedback may be at least one that the user 101 has “fully accepted”, “partially accepted” or “not accepted” the given executable source code 109. In case, the user 101 has “fully accepted” the given executable source code 109, the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106. Also, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements. On the other hand, if the given executable source code 109 is “partially accepted” by the user 101, then the partially accepted part of the source code may be passed to the pre-trained code generation model 106 for the reinforcement learning and to learn the mistake. Also, a penalty may be assigned to the pre-trained code generation model 106, so that the pre-trained code generation model 106 ensures not to retrieve same or similar executable source codes 109 for the similar application requirements. Similar learning strategy is followed when the user feedback is “not accepted”.).
With respect to claim 18, Kozhisseri also discloses receiving additional code data that includes the code block (e.g., Figs. 1-4 and associated text, e.g., col. 13:3-23, In case, the user 101 has “fully accepted” the given executable source code 109 [additional code data that includes the code block], the selected executable source code 109 may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106.; col. 15:30-24, As indicated in block 411 of the method 400, the code generation model 106 of the source code generator 105 may undergo a reinforcement learning and/or relearning based on the user feedback on the executable source code 109.);
scanning the additional code data (Id., particularly, the selected executable source code 109 [additional code block] may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106.); and
generating an additional NLP model based on at least a portion of the code data and at least a portion of the additional code data (Id.; col. 15:3-8, the at least one pre-trained code generation model 106 [NLP model] may be trained by classifying the one or more reference source codes, retrieved from the predetermined repositories [the code data]; col. 13:3-23, In case, the user 101 has “fully accepted” the given executable source code 109, the selected executable source code 109 [additional code data] may be fed back to the pre-trained code generation model 106 as an input to reinforcement learning of the pre-trained code generation model 106. Also, the pre-trained code generation model 106 may consider this instance as a “reward” and learn to recommend similar executable source codes 109 for similar application requirements. On the other hand, if the given executable source code 109 [additional code data] is “partially accepted” by the user 101, then the partially accepted part of the source code may be passed to the pre-trained code generation model 106 for the reinforcement learning and to learn the mistake. Also, a penalty may be assigned to the pre-trained code generation model 106, so that the pre-trained code generation model 106 ensures not to retrieve same or similar executable source codes 109 for the similar application requirements [code generation model 106 is retrained using the generated executable source code 109, i.e., additional code data, thereby creating a changed code generation model, i.e., generating an additional NLP model based on at least a portion of the additional code data, and the executable source code 109 used for the retraining was generated by the code generation model 106, which was pre-trained on the code data retrieved from the repositories, i.e., generating an additional NLP model based on at least a portion of the code data]; see also col. 15:30-24.).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Specifically, Sanh et al., “DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter” discloses pre-training a smaller NLP model based on a larger NLP model.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to STEPHEN DAVID BERMAN whose telephone number is (571) 272-7206. The examiner can normally be reached M-F, 9-6 Eastern.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Hyung S. Sough can be reached on 571-272-6799. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/STEPHEN D BERMAN/ Examiner, Art Unit 2192
Prosecution Insights