Projects and Architecture

Spring 2024 |

HR-VQLoRA: Using Hierarchical Adaptors to Fine Tune LLMs

A better way to fine-tune LLMs on consumer-grade hardware.

    Quantize LLMs from 32-bit to 4-bit
+ Training a low-rank (small) ΔW matrix
+ Adding a hierarchical adaptor
+ Including a custom loss function
+ Integrating into HuggingFace PEFT and Transformers

    = One line integrations
    = 8% better performance on MMLU

Read More.

Podlite: Your daily aggregate podcast

Combining similar clips from your favorite podcasts with AI

HackDartmouth 2024 Runner Up
Read More.

Conf-ViT: A Violence Early-Warning System

SOTA conflict forcasting at a subnational level using the Transformer

Read More.

Winter 2024 |

Discite: A Learning Platform for the Modern Attention Span

Lorem ipsum dolor sit amet.

Spring 2023 |

AI Explainability in Self Driving

A whitepaper on explainable AI.

Spring 2021 |

Spyfall Offline

an offline game to play when camping with 77k+ downloads