Method and system for multi-level artificial intelligence supercomputer design
A system and method for enhancing performance of h-LLMs including providing specialized h-LLMs specialized for different tasks and trained on a training dataset specific to the specific task to be performed by the specialized h-LLM, generating synthetic data by processing input prompts through the h-LLMs to produce synthetic outputs, feeding the synthetic data back to the specialize h-LLMs through a feedback loop process implemented in a cloud container environment, and performing model refinement on the specialized h-LLMs by retraining at least one specialized h-LLM using the synthetic data.
1 . A method for enhancing performance of one or more families of large language models (h-LLMs) comprising:
providing a plurality of specialized h-LLMs, each specialized h-LLM of the plurality of specialized h-LLMs being specialized to perform different specific tasks, each specialized h-LLM of the plurality of specialized h-LLMs being a family of large language models trained on a training dataset specific to the specific task to be performed by the specialized h-LLM;
generating synthetic data by processing input prompts through the plurality of specialized h-LLMs to produce synthetic outputs;
feeding the synthetic data back to the plurality of specialized h-LLMs through a feedback loop process implemented in a cloud container environment; and
performing model refinement on the plurality of specialized h-LLMs by retraining at least one specialized h-LLM of the plurality of specialized h-LLMs using the synthetic data.
2 . The method of claim 1 further comprising:
generating a base h-LLM model by performing unsupervised pre-training on input data; and
generating the plurality of specialized h-LLMs by performing supervised fine-tuning on the base h-LLM model.
3 . The method of claim 1 wherein the specific tasks are selected from the group consisting of question answering, information extraction, sentiment analysis, image captioning, object recognition, instruction following, classification, inferencing, and sentence similarity.
4 . The method of claim 1 wherein the step of performing model refinement comprises a bagging technique comprising:
partitioning input data into a plurality of distinct subsets;
training a plurality of preliminary h-LLMs on different subsets of the plurality of distinct subsets having lower individual precision and accuracy in parallel; and
merging or fusing the plurality of preliminary h-LLMs to create a merged h-LLM with higher precision and accuracy than each preliminary h-LLM of the plurality of preliminary h-LLMs.
5 . The method of claim 1 wherein the step of performing model refinement comprises a boosting technique comprising:
training a first h-LLM using original input data;
evaluating an output of the first h-LLM;
generating weighted data by assigning weights to the output based on the evaluation of the output;
training a second h-LLM using the weighted data;
sequentially creating subsequent h-LLMs by iteratively evaluating output of the second h-LLM and subsequent h-LLMs, generating subsequent weighted data, and training subsequent h-LLMs; and
merging the first h-LLM, second h-LLM, and all subsequent h-LLMs to create a merged h-LLM.
6 . The method of claim 1 wherein the step of performing model refinement comprises an extraction technique comprising:
performing an extraction and specialization process on a large general-purpose h-LLM to extract specific knowledge required for a specific task; and
creating a smaller h-LLM specialized for the specific task, wherein the smaller h-LLM is more efficient for the specific task than the large general-purpose h-LLM.
7 . The method of claim 1 further comprising:
training a first h-LLM using text data;
training a second h-LLM using image data;
training a third h-LLM using audio data; and
combining the first h-LLM, second h-LLM, and third h-LLM through a merging process to create a merged multimodal h-LLM capable of processing text, image, and audio inputs.
8 . The method of claim 1 wherein the step of performing model refinement on the plurality of specialized h-LLMs cause the specialized h-LLMs of the plurality of specialized h-LLMs to have at least one of improved accuracy and improved performance metrics.
9 . A system for answering queries using retrieval-augmented generation comprising:
a processor;
a communication device positioned in communication with the processor; and
a non-transitory computer-readable storage medium positioned in communication with the processor and having stored thereon software executable by the processor to instantiate:
a plurality of embedding models operable to convert user prompts into prompt embeddings;
a vector database storing knowledge documents and operable to return relevant knowledge documents based on similarity of embeddings to the user prompts;
an AI input broker operable to create context-aware prompts based on the user prompts and the relevant knowledge documents; and
a plurality of families of large language models (h-LLMs) operable to process the context-aware prompts and generate results.
10 . The system of claim 9 wherein:
the vector database further stores knowledge document embeddings, each knowledge document embedding being associated with a knowledge document stored on the vector database; and
the vector database is further operable to determine similarity between the prompt embeddings and knowledge document embeddings to identify the relevant knowledge documents.
11 . The system of claim 9 wherein the plurality of h-LLMs are operable to provide citations from the relevant knowledge documents along with the results to indicate source references for the generated results.
12 . A system for creating enhanced families of large language models (h-LLMs) comprising:
a processor;
a plurality of h-LLMs having different precision and accuracy levels;
a non-transitory computer-readable storage medium having software stored thereon that, when executed by the processor, is operable to:
apply a bagging technique by merging at least a first subset of the plurality of h-LLMs trained on different data subsets to create a merged h-LLM with higher accuracy;
apply a boosting technique by sequentially training at least a second subset of the plurality of h-LLMs with weighted data to create progressively better models; and
apply an extraction technique by creating one or more smaller specialized h-LLMs from one or more larger general-purpose h-LLMs.
13 . The system of claim 12 wherein the one or more smaller specialized h-LLMs are optimized for specific tasks selected from the group consisting of question answering, information extraction, sentiment analysis, image captioning, object recognition, instruction following, classification, inferencing, and sentence similarity.
14 . The system of claim 12 wherein the software is further operable to:
generate synthetic data using output from the plurality of h-LLMs;
feed the synthetic data back to the plurality of h-LLMs through a feedback loop; and
perform iterative model refinement to continuously improve performance of at least one h-LLM of the plurality of h-LLMs.
15 . The system of claim 12 wherein the bagging technique comprises:
partitioning training data stored on the storage medium into a plurality of distinct data subsets;
training the first subset of the plurality of h-LLMs on the plurality of distinct data subsets in parallel; and
merging the first subset of the plurality of h-LLMs to create the merged h-LLM.