|
|
--- |
|
|
title: README |
|
|
emoji: 🏃 |
|
|
colorFrom: indigo |
|
|
colorTo: gray |
|
|
sdk: static |
|
|
pinned: false |
|
|
thumbnail: >- |
|
|
https://cdn-uploads.huggingface.co/production/uploads/65e5fd212faa026716fd27bf/P8yHyv6OZSPLpYEPE0XII.png |
|
|
--- |
|
|
# Welcome to GEM Space 🫡 |
|
|
Greetings from GEM Space, the heart of innovation behind our paper, "FRAGILE MASTERY: ARE DOMAIN-SPECIFIC TRADE-OFFS UNDERMINING ON-DEVICE LANGUAGE MODELS?". We’re thrilled to invite you into our world of edge AI exploration! This repository, GEM_Testing_Arsenal, is a cornerstone of our efforts to redefine On-Device Language Models (ODLMs) through the Generalized Edge Model (GEM). Keep an eye out for the paper link once it’s published! |
|
|
|
|
|
> Link to the Paper: [Click Here](https://arxiv.org/abs/2503.22698) |
|
|
--- |
|
|
## About Our Paper 📄 |
|
|
|
|
|
***Abstract***: |
|
|
|
|
|
The deployment of On-Device Language Models (ODLMs) on resource-constrained edge devices demands a delicate balance of efficiency, memory, power, and linguistic skill across diverse tasks. In "FRAGILE MASTERY", we explore the trade-offs between domain-specific optimization and cross-domain robustness, introducing the Generalized Edge Model (GEM). GEM integrates specialization and generalization using a Sparse Cross-Attention Router (SCAR), achieving a cross-domain F1 score of 0.89 with sub-100ms latency on platforms like Raspberry Pi 4 and Pixel 6. Across 47 benchmarks spanning eight domains—healthcare, legal, finance, STEM, and more—GEM boosts general-task performance by 7% over GPT-4 Lite while matching domain-specific results. With new metrics like the Domain Specialization Index (DSI) and a balanced distillation framework cutting catastrophic forgetting by 43%, this work offers a robust foundation for edge AI. [Paper Link](https://arxiv.org/abs/2503.22698) |
|
|
|
|
|
***Architecture***: |
|
|
<div align="center"> |
|
|
<img src="https://huggingface.co/datasets/GEM025/GEM_Arsenal/resolve/main/GEM.png" width="500px"> |
|
|
<p><b>Fig: </b><i>Our Architecture simplified</i></p> |
|
|
</div> |
|
|
|
|
|
***Our Vision***: |
|
|
|
|
|
At GEM Space, we’re on a mission to revolutionize edge intelligence. We’re striving to build On-Device Language Models that: |
|
|
- **Thrive Under Constraints**: Deliver exceptional accuracy and speed on low-power devices—from smartphones to custom NPUs—without compromise. |
|
|
- **Master the Balance**: Seamlessly blend domain-specific expertise (think healthcare diagnostics or financial analysis) with robust, cross-domain adaptability. |
|
|
- **Empower the Edge**: Bring advanced AI to the fingertips of real-world applications, making it fast, practical, and accessible wherever it’s needed. |
|
|
|
|
|
The [**GEM_Testing_Arsenal**](https://huggingface.co/datasets/GEM025/GEM_Arsenal) embodies this ambition—a testing ground for GEM, our pioneering architecture designed to make ODLMs smarter, leaner, and more versatile. We’re here to push the limits of what’s possible and inspire a new era of edge AI innovation. |
|
|
|
