Team - Sam Brown, Basil Labib, Codruta LugojI worked on building a meta-benchmark to evaluate how well agents can improve other agents on consequential benchmarks. Specifically, I worked on the WMDP benchmark which evaluates models for its dangerous knowledge related to cybersec, chemical security, and biosecurity. I developed a task that evaluates if an LLM agent can perform unlearning of WMDP knowledge on a Llama-3-8B model. I have also done the unlearning myself. Currently evaluating the… Show more Team - Sam Brown, Basil Labib, Codruta LugojI worked on building a meta-benchmark to evaluate how well agents can improve other agents on consequential benchmarks. Specifically, I worked on the WMDP benchmark which evaluates models for its dangerous knowledge related to cybersec, chemical security, and biosecurity. I developed a task that evaluates if an LLM agent can perform unlearning of WMDP knowledge on a Llama-3-8B model. I have also done the unlearning myself. Currently evaluating the agents, addressing their spurious failure modes, and writing a paper. Show less

Worked under the guidance of Dr. Jayashree R at PES University. Tackled a problem of document classification on news articles in Kannada using various word representations.

As a student mentor at PESU I/O, I had the privilege to teach Machine Learning with TensorFlow to a group of juniors,

During my time at Datacogin, I have worked with various tools and technologies like Python, Flask, MongoDB interfaces for Python, Keras and Tensorflow. My work included developing REST APIs and training ML models for image classification.