Worked on preference modelling for LLM alignment. Conducted experiments to investigate the relationship between quality, quantity and annotation acquisition method of preference data on performance of models trained with different alignment techniques (DPO, cDPO, IPO). Published the work in NAACL Workshop.

- Contingent leader for the Inter IIT Tech Meet 12.0 in which IIT Patna finished 8th among all IITs - Established Research Community and new clubs/chapters - Revived the Grand Championship - a semester-long inter-house tech competition featuring problem statements, workshops from different tech clubs

Worked on instruction-based finetuning of large language models on conversation data. - Generated training data with a self-instruct pipeline with diversity in prompting techniques and tasks- Created a leaderboard on instruction prompts and fine-tuned the best-performing model on the augmented dataShowed that instruction-based fine-tuning helps with the instruction alignment and robustness of the model on conversational data.Offered a full-time role

- Helped establish a separate, independent community for ML enthusiasts under NJACK, the official computer science club- Developed roadmaps and conducted introductory sessions on ML theory and application for the juniors - Organized and held talks with research professor and paper reading sessions

- Revamped the website of SAAR with an enhanced aesthetic using modern front-end frameworks- Opensourced the project during Codepeak and mentored interested individuals