Quantiphi, Inc

A system ( 100 ) for fine-tuning a code translation model is disclosed. The system comprises a processing arrangement ( 102 ) and a user device ( 104 ). The processing arrangement is configured to receive at least one seed sample in a source code language, a code translation base model, and a set of system specific requirements, from a user device. The processing arrangement is configured to generate synthetic dataset comprising training dataset and evaluation dataset; determine at least one parameter to finetune the code translation model; finetune the code translation model based on at least one parameter; evaluate the code translation model to generate a evaluation score; determine a training loss and an evaluation loss; based on the evaluation score, the training loss and the evaluation loss, finetune the code translation model or approve the code translation model. A method for fine-tuning a code translation model is also disclosed.

Disclosed is method for distributed decision-making in multi-role large language models (LLMs) architecture (mLLMa). Method comprises: receiving user query for initiating conversation between user and mLLMa; using graphical representation (GR) for identifying role(s) associated with user query, wherein role(s) is based on context thereof; assigning relevance score to each role; dynamically passing role(s), to mLLMa; generating role-specific prompt for role assumed by LLM; generating, by each LLM, role-specific response (RR) corresponding to role-specific prompt for each role and context; presenting RR from each LLM to peer LLMs; conducting polling process among LLMs for ranking RR therefrom; aggregating rankings from polling process to determine final ranking of RR for each LLM; selecting RR having highest final ranking as an action-inducing response (AR), and transmitting AR to user for providing user action; and updating GR based on AR and user action.

Disclosed is method for generating a target molecule ( 412, 514, 802 ), comprising receiving first user input ( 806 ) indicative of properties associated with target molecule, and identifying properties ( 808 A-C) associated with targeted molecule and corresponding objectives ( 810 A-B); generating property scores ( 832 A-B) for properties using property predictor algorithm ( 812 ); receiving second user input indicative of molecular structure ( 202, 402, 502, 602, 702, 814 ) of input molecule ( 204 ); generating corresponding target molecules (CTMs) ( 200, 406, 508, 600, 700, 816 ); generating embeddings ( 824 ) of CTMs; determining aggregate similarity score ( 828 ); determining aggregate property score; determining fitness scores ( 410, 512, 834 ) of CTMs; determining whether given target molecule amongst CTMs fulfill termination criteria (TC); when it is determined that TC is fulfilled by given target molecule, deeming given target molecule as target molecule to be generated; when it is determined that TC is not fulfilled, updating generated CTMs, iteratively performing steps (v) to (ix).

A method and system for multi-table question answering is provided herein. The method includes receiving a complex question from a user. The method may further include decomposing the complex question into one or more sub-questions using one or more of but not limited to coreference resolution, dependency parsing, chunking and prompt techniques. The method may further include generating code scripts for each sub-question using a Large Language Model (LLM). The method may further include executing the code scripts to retrieve relevant structured data from a plurality of tables. The method may further include generating an answer to the complex question based on the retrieved structured data.

The present disclosure provides a method for recommendation of suitable assets using Large Language Model (LLM), method comprising receiving descriptors for each asset from amongst plurality of assets; generating embeddings for each asset from amongst plurality of assets, by employing LLM; creating database of assets by storing generated embeddings for each asset from amongst plurality of assets in database; receiving user query pertaining to request for recommendation to solve enterprise problem; generating cosine similarity scores, wherein each cosine similarity score is generated between user query and corresponding generated embedding for given asset stored in database of assets; identifying predefined number of similar assets from amongst database of assets, based on highest cosine similarity scores; constructing optimal prompt based on identified predefined number of similar assets and user query; and prompting LLM, using constructed optimal prompt for generating response as recommendation of identified predefined number of similar assets as suitable assets for solving user query.

A method for electronic processing of user queries maintaining factual consistency during processing includes receiving a user query, executing a hybrid retrieval operation, generating a context information space that includes a set of paragraph output from the paragraph retrieval operation, a set of triplet output from the triplet retrieval operation, and a set of graphs output from the graph retrieval operation. Furthermore, generating a first prompt instruction to generate an intermediate rationale as explanation output and performing an automatic fact-checking operation to verify the explanation output and re-feeding the user query along with the generated explanation output to the generative language model to derive an answer output that is succinct in one or more language parameters as compared to the generated explanation output. The answer output includes a verified and evidence-supported answer and at least one fact reference from first set of factual information with reduced hallucinations and factual inconsistencies.

A method and system for developing a generative chemistry model based on SAFE representations is disclosed. The method includes encoding one or more chemical compounds into the SAFE representations. The method may include training an encoder-decoder transformer model based on the SAFE representations using one or more masking techniques. The encoder-decoder transformer model creates a latent space to represent encoded SAFE representations of the chemical compounds. The method may further include training a variational Autoencoder (VAE) to generate a continuous latent space by compressing the latent space of the encoder-decoder transformer model. The method may further include optimizing the continuous latent space to generate a plurality of chemical compounds with specific properties by decoding the SAFE representations of the chemical compounds.

A method and system for multi-table question answering is provided herein. The method includes receiving a complex question from a user. The method may further include decomposing the complex question into one or more sub-questions using one or more of but not limited to coreference resolution, dependency parsing, chunking and prompt techniques. The method may further include generating code scripts for each sub-question using a Large Language Model (LLM). The method may further include executing the code scripts to retrieve relevant structured data from a plurality of tables. The method may further include generating an answer to the complex question based on the retrieved structured data.