Sonu Dixit

Industry Experience

LLama 3 Finetuing for Customer Chat Conversation

• We finetune LLama 3.1 8B model on Agent Customer conversation data.
• After fine-tuning, the model shifted from generating verbose responses to concise, agent-like replies.
• It began personalizing conversations by using customer names and adopting an agent identity, which was absent before fine-tuning. This effect might also be achievable through prompt engineering.
• Performance significantly improved in initial and final conversation turns.
• However, the model struggled to generate relevant responses during middle turns that required external information.
• Performance on instruction-following benchmarks declined, likely due to the lack of diversity in training instructions. To address this, I recommend using a mixed dataset—70% chat data and 30% diverse tasks (e.g., instruction-following, reasoning, math, etc.).

Intent Classification

• We classify customer queries into predefined intents to enable automated handling by our Chatbot. Our trained model outperforms the existing solution and LLMs including GPT-3.5.
• Transformer Based multiple network(s) training.
• Classes are added/removed with time. Final solution design must take this into account.
• We find manual finetuning to be better than data efficient methods like SetFit
• We compare against llms (GPT, Mixtral, and others) using few shot prompt-tuning and finetuning(generation based classification)
• We design and conduct experiments to estabilish the approach can understand client-useful words even when its OOV for the model.
• We conduct experimens to study its performance on OOD data, and how to minimise the risk.
• We found bias in models behaviour. We design solutions to handle such bias.

Information Retrieval and Conditional Generation

• Trained a DistilBERT on the MSMarco dataset to proactively recommend relevant articles, achieving comparable performance on the TREC21 and TREC20 datasets.
• We train multiple formulations Bert_dot, Bert_cat, ColBert for ranking task. Bert_cat outperforms other formulations.
• Bert_cat requires a lot of inference time compute. To reduce latency We perform a dual distil from ensemble scores of Bert_cat and ColBert to the DistilBERT
• The method is inspired from Efficiently teaching an effective dense retriever with balanced topic aware sampling
• We explore multiple first level rankers (TF-IDF, BM25, LSH, PQ-Encoding)
• We finetune the trained DistilBERT on our client specific data.
• Recommendation from a database of Question-Answer pairs. Upon finetuning, We outperform the exisiting solution.
• Conditional response generation for an ongoing conversation. Results on public data team_name-Test_Team_Name
• End to end training of RAG system using posterior guided retriever. Here we use the trained DBert as the retriever.
• We manually evaluate the trained system on a client specific data (web-scraped client website data and questions)
• We worked on ATLAS: Few-shot Learning with Retrieval Augmented Language Models

LLM as Co-Pilot

• We utilize LLMs to generate contextually relevant responses during conversations.
• We append the LLM context with relevant historical information, using the retriever trained in the IR task.
• We designed an identity task (similar to Needle in a Haystack) to evaluate the LLM performance.

Compliance Evaluation and Early Warning System - Agent-Customer Chat

• We pass the quantifiable metrics to LLM.
• Prompt-Engineering, We use LLM-as-judge for evaluation.

Named Entity Recognition by Question Answering

• Different clients need different granular entities(start_date_of_reservation, flight_time, ..). Traditional NER requires training data for the specific entities.
• We formulate this as QA problem, where question is asked for client specific entities. Model needs to Generate/extract the answer(entity). The same underlying QA-system can generalise to multiple clients.
• It reduces client-specific entity extraction finetuning.

Content Analysis and Storyline Extraction

• Quantified character attributes using text, image, and audio data, correlating features with viewership metrics.
• Episode vector learning, social relationship, face recognition. Website

Entity extraction from Semi-Structured Documents

• Used HMM to extract values corresponding to predefined keys from documents.
• Implemented the Viterbi and the scaled Baum-Welch algorithm for learning parameters.

Publications

Zero-Shot Generalization using Intrinsically Motivated Compositional Emergent Protocols

Authors: R Hazra*, Sonu Dixit* — NAACL 2021 Workshop (ViGIL)

Latest Preprint:Intrinsically Motivated Compositional Language Emergence

We argue intrinsic rewards increase compositionality in emergent communication between two Agents. The improved compositionality increases zero shot generalisation in the downstream tasks. website

gComm: Environment for Investigating Generalization in Grounded Language Acquisition

Authors: R Hazra, Sonu Dixit, S Sen — NAACL 2021 Workshop (ViGIL)

This work introduces gComm, an environment designed to evaluate grounded language acquisition. We focus on how agents generalize language understanding in complex, multi-agent scenarios.

M.Tech Thesis: Adaptive Traffic Signal Control using Multi Agent RL

• Implemented multi-agent reinforcement learning (MARL) to dynamically adjust traffic signal durations based on congestion.
• Algorithm: Proximal Policy Optimization (PPO) with Advantage Actor-Critic. Report
• Simulated traffic data using PTV Vissim and demonstrated MARL’s superiority over fixed-time algorithms in terms of average speed, delay, and lane occupancy. Results
• Each Signal is represented as a Neural Network. During training multiple small neural networks are trained at same time. The networks are connected via the reward that they get.
• Advisor - Prof Shalabh Bhatnagar

Contact

Email: sonudixit2k@gmail.com

LinkedIn: linkedin.com/in/sonudixit