Hi there! My name is Srishti. I am a Machine Learning Engineer and Researcher, and I care the most about the following areas of study:
- Alignment of Language Models (For example: While, in theory, SFT teaches the models to lie and RLHF helps get around this by virtue of Reward Models, how well do RMs work specifically for this purpose?)
- Techniques for efficient training of and inference from Language Models (Here, I am particularly interested in writing software that makes the most efficient use of the hardware at hand.)
These days I am focussing a lot on practicing DS & Algos on Leetcode and studying how Operating systems work. The latter is so fun!
Work
Started working with Aran Komatsuzaki towards building a high quality training+eval dataset for PRMs (Process Reward Models).
I am fortunate to be working with Alex Havrilla towards an attempt to train via RL a potentially new SOTA Math LLM. :)
I am leading a group at Cohere for AI, Expedition Aya where we are conducting experiments aimed at better understanding of some open source Reward Models and of some proprietary LLMs-as-Judges in a multilingual setting as far as their ability to align with human preferences is concerned.
My full time job is as an ML Engineer at an Indian startup where I work on many things ML + software engineering. An average day involves building and/or debugging backend ML pipelines from scratch. Sometimes I also build new features for existing products.
In the past, I have worked at Translated where I had a lot of fun working on the intersection of LLMs, DP, FL, PEFT for the EU Data Tools for Heart Project.
I was also one of the 10 fellows that were selected to be a part of Pi School where I worked on using many traditional and neural NLP techniques to automate an end to end document processing and analysis pipeline for a Berlin-based startup, Briink.
I started teaching myself Machine learning in late 2021 using online courses. In the Jan of 2022, I worked on evaluating the robutness of BERT based models for the task of biomedical entity linking - we got this published as a workshop paper @ NeurIPS ‘22.
Jan 2022 is also the time when I got deeply interested in the internal workings of PyTorch which led me to studying and exploring more of the library an outcome of which is something I cherish - I was awarded by The Linux Foundation and The PyTorch Foundation with one of the 12 PyTorch Contributor Awards 2023.