核心概念
Promoting completeness and openness in AI models is crucial for transparency, reproducibility, and innovation.
要約
The Model Openness Framework (MOF) proposes a ranked classification system to rate machine learning models based on their completeness and openness. It aims to prevent misrepresentation of open models, guide researchers in providing all components under open licenses, and foster a more open AI ecosystem. The framework consists of three classes - Open Model, Open Tooling, and Open Science - each representing different levels of completeness and openness.
Introduction
AI advancements driven by computational capabilities.
Concerns about transparency, reproducibility, ethics, and safety.
Debate on the benefits and risks of open models.
Challenges with "Open" Models
Lack of transparency in state-of-the-art foundation models.
Misleading claims of being "open-source."
Issues with restrictive licenses leading to "open-washing."
Importance of Completeness and Openness
Need for releasing all artifacts for transparency.
Differentiating between completeness and openness.
Importance of open licenses for data parameters.
Value of Openness in AI
Power of open source software.
Benefits of open data sharing.
Promise of open access publications.
MOF Components
Datasets
Data Preprocessing Code
Model Architecture
Model Parameters
Model Metadata
Training, Validation & Testing Code
Inference Code
Evaluation Code
Evaluation Data
統計
Many "open-source" GAI models lack necessary components for full understanding.
Some use restrictive licenses known as "openwashing."
引用
"Openness is not just about what is included but importantly under what licenses each component is released." - MOF Proposal