Method and system for evaluating integration of responsible AI with LLM operations
A computer-implemented method for evaluating integration of Responsible Artificial Intelligence Operations (RAIOPS) and Large Language Model Operations (LLMOPS) is disclosed. A response respective to each of prompts is generated using an LLM, in response to receiving data associated with each of the prompts. The data associated with each of the prompts and data associated with the response respective to each of the prompts is stored as an association. Further, based on user-specified criteria and using the data associated with the prompts or the data associated with the responses respective to the prompts, one or more evaluation metrics are generated for evaluating the responses respective to each of the prompts for one or more aspects. In accordance with the at least one evaluation metric, a knowledge graph visualization or a numerical score is generated to display performance of the LLM and determine whether the LLM needs optimization or tuning.
1 . A computer-implemented method for evaluating integration of Responsible Artificial Intelligence Operations (RAIOPS) and Large Language Model Operations (LLMOPS), comprising:
generating, by one or more processors, in response to receiving data associated with each prompt of a plurality of prompts, a response respective to each prompt of the plurality of prompts using at least one Large Language Model (LLM);
storing, by the one or more processors, in at least one memory, the data associated with each prompt of the plurality of prompts and data associated with the response respective to each prompt as an association;
generating, by the one or more processors, based at least in part upon a user-specified criteria and using the data associated with a subset of the plurality of prompts or the data associated with the response respective to the subset of the plurality of prompts, at least one evaluation metric for evaluating the response respective to each prompt of the plurality of prompts for at least one aspect of a plurality of aspects;
generating, by the one or more processors, in accordance with the at least one evaluation metric, a knowledge graph visualization or a numerical score to display performance of the LLM and determine whether the LLM needs optimization or tuning; and
boosting, by the one or more processors, the numerical score upon finding a synonym match in the response when compared with a respective ground-truth, and wherein the synonym match utilizes a Bidirectional Encoder Representations from Transformers (BERT) multilingual model,
wherein the boosting comprises:
calculating metrics specific to a domain or an application for the data associated with each prompt of the plurality of prompts and the data associated with the response respective to each prompt of the plurality of prompts,
analyzing central tendencies of the metrics derived from multiple runs to generate different responses,
setting a baseline score of each metric based on the analyzed central tendencies, and
calculating a safe score based on the baseline score to determine the boosting of the numerical score.
2 . The computer-implemented method of claim 1 , wherein the user-specified criteria include generating the at least one evaluation metric at a preconfigured time interval and/or generating the at least one evaluation metric upon generating a preconfigured number of responses.
3 . The computer-implemented method of claim 1 , wherein the plurality of aspects includes relevance, inconsistency, security, drift detection, robustness, bias and fairness detection, accuracy and appropriateness of the response, transparency and explainability, hallucination detection, and/or language translation or caching sustainability.
4 . The computer-implemented method of claim 1 , further comprising prior to generating the at least one evaluation metric, performing, by the one or more processors, dimensionality reduction techniques or clustering techniques on the data associated with the subset of the plurality of prompts and/or the data associated with the response respective to the subset of the plurality of prompts.
5 . The computer-implemented method of claim 1 , wherein the at least one evaluation metric generated for drift detection identifies a content drift, a data drift, a temporal drift, a tone drift, an upstream drift, a domain drift, a covariate drift, a prior probability drift, a population drift, a feature drift, a sampling bias drift, a seasonal drift, a conceptual drift, an adversarial attach drift, an environmental drift, a response drift, a prompt drift, and/or embeddings drift.
6 . The computer-implemented method of claim 1 , wherein the at least one evaluation metric generated for relevance evaluates the response for at least one of misinformation, abuse, toxic content, bias, text inconsistencies, and/or relevancy.
7 . The computer-implemented method of claim 1 , wherein the at least one evaluation metric generated for security evaluates the subset of the plurality of prompts for a prompt injection attack, a prompt leakage attack, a prompt poisoning attack, and/or a prompt jailbreaking attempt.
8 . The computer-implemented method of claim 1 , further comprising generating, by the one or more processors, a plurality of selections to provide for optimization and/or tuning of the LLM.
9 . The computer-implemented method of claim 1 , further comprising mitigating hallucinations in the response respective to each prompt of the plurality of prompts generated by the LLM by:
retrieving, from a vector database, top-k matches corresponding to the response respective to each prompt of the plurality of prompts generated by the LLM, and
comparing the response respective to each prompt of the plurality of prompts generated by the LLM with the top-k matches to assess alignment, wherein the comparison includes calculating one or more statistical metrics for generating similarity scores to mitigate the hallucinations in the response respective to each prompt of the plurality of prompts.
10 . The computer-implemented method of claim 1 , further comprising evaluating transparency of the response respective to each prompt of the plurality of prompts by:
determining general statistics based on pre-processed data of the LLM, and
determining transparency through analysis of a 360-degree view of the data associated with each prompt of the plurality of prompts and the data associated with the response respective to each prompt of the plurality of prompts and
integrating transparency measures to assess distribution and complexity of the response respective to each prompt of the plurality of prompts to evaluate the transparency of the response respective to each prompt of the plurality of prompts.
11 . A system for evaluating integration of Responsible Artificial Intelligence Operations (RAIOPS) and Large Language Model Operations (LLMOPS), the system comprising:
at least one memory storing machine-executable instructions; and
at least one processor communicatively coupled with the at least one memory, wherein the at least one processor executes the machine-executable instructions to perform operations comprising:
generating, in response to receiving data associated with each prompt of a plurality of prompts, a response respective to each prompt of the plurality of prompts using at least one large language model (LLM);
storing, in the at least one memory, the data associated with each prompt of the plurality of prompts and data associated with the response respective to each prompt as an association;
generating, based at least in part upon a user-specified criteria and using the data associated with a subset of the plurality of prompts or the data associated with the response respective to the subset of the plurality of prompts, at least one evaluation metric for evaluating the response respective to each prompt of the plurality of prompts for at least one aspect of a plurality of aspects; generating, in accordance with the at least one evaluation metric, a knowledge graph visualization or a numerical score to display how the LLM is performing and for a user to determine whether the LLM needs optimization or tuning; and
boosting the numerical score upon finding a synonym match in the response when compared with a respective ground-truth, and wherein the synonym match utilizes a Bidirectional Encoder Representations from Transformers (BERT) multilingual model,
wherein the boosting comprises:
calculating metrics specific to a domain or an application for the data associated with each prompt of the plurality of prompts and the data associated with the response respective to each prompt of the plurality of prompts,
analyzing central tendencies of the metrics derived from multiple runs to generate different responses,
setting a baseline score of each metric based on the analyzed central tendencies, and
calculating a safe score based on the baseline score to determine the boosting of the numerical score.
12 . The system of claim 11 , wherein the user-specified criteria include generating the at least one evaluation metric at a preconfigured time interval and/or generating the at least one evaluation metric upon generating a preconfigured number of responses.
13 . The system of claim 11 , wherein the plurality of aspects includes relevance, inconsistency, security, drift detection, robustness, bias and fairness detection, accuracy and appropriateness of the response, transparency and explainability, hallucination detection, and/or language translation or caching sustainability.
14 . The system of claim 11 , wherein the operations further comprise prior to generating the at least one evaluation metric, performing dimensionality reduction techniques or clustering techniques on the data associated with the subset of the plurality of prompts and/or the data associated with the response respective to the subset of the plurality of prompts.
15 . The system of claim 11 , wherein the at least one evaluation metric generated for drift detection identifies a content drift, a data drift, a temporal drift, a tone drift, an upstream drift, a domain drift, a covariate drift, a prior probability drift, a population drift, a feature drift, a sampling bias drift, a seasonal drift, a conceptual drift, an adversarial attach drift, an environmental drift, a response drift, a prompt drift, and/or embeddings drift.
16 . The system of claim 11 , wherein the at least one evaluation metric generated for relevance evaluates the response for at least one of misinformation, abuse, toxic content, bias, text inconsistencies, and/or relevancy.
17 . The system of claim 11 , wherein the at least one evaluation metric generated for security evaluates the subset of the plurality of prompts for a prompt injection attack, a prompt leakage attack, a prompt poisoning attack, and/or a prompt jailbreaking attempt.
18 . The system of claim 11 , wherein the operations further comprise generating a plurality of selections to provide for optimization and/or tuning of the LLM.
19 . The system of claim 11 , wherein the operations further comprise mitigating hallucinations in the response respective to each prompt of the plurality of prompts generated by the LLM by:
retrieving, from a vector database, top-k matches corresponding to the response respective to each prompt of the plurality of prompts generated by the LLM, and
comparing the responses respective to each prompt of the plurality of prompts generated by the LLM with the top-k matches to assess alignment, wherein the comparison includes calculating one or more statistical metrics for generating similarity scores to mitigate the hallucinations in the response respective to each prompt of the plurality of prompts.
20 . A non-transitory computer-readable media (CRM) comprising instructions stored thereon for evaluating integration of Responsible Artificial Intelligence Operations (RAIOPS) and Large Language Model Operations (LLMOPS), wherein the instructions, when executed by at least one processor of a computing device, cause the computing device to perform operations comprising:
generating, in response to receiving data associated with each prompt of a plurality of prompts, a response respective to each prompt of the plurality of prompts using at least one large language model (LLM);
storing, in at least one memory, the data associated with each prompt of the plurality of prompts and data associated with the response respective to each prompt as an association;
generating, based at least in part upon a user-specified criteria and using the data associated with a subset of the plurality of prompts or the data associated with the response respective to the subset of the plurality of prompts, at least one evaluation metric for evaluating the response respective to each prompt of the plurality of prompts for at least one aspect of a plurality of aspects;
generating, in accordance with the at least one evaluation metric, a knowledge graph visualization or a numerical score to display how the LLM is performing and for a user to determine whether the LLM needs optimization or tuning; and
boosting the numerical score upon finding a synonym match in the response when compared with a respective ground-truth, and wherein the synonym match utilizes a Bidirectional Encoder Representations from Transformers (BERT) multilingual model,
wherein the boosting comprises:
calculating metrics specific to a domain or an application for the data associated with each prompt of the plurality of prompts and the data associated with the response respective to each prompt of the plurality of prompts,
analyzing central tendencies of the metrics derived from multiple runs to generate different responses,
setting a baseline score of each metric based on the analyzed central tendencies, and
calculating a safe score based on the baseline score to determine the boosting of the numerical score.