watsonx.ai

ย View Only

Introducing InstructLab - A new community-based approach to build truly open-source LLMs

By Armand Ruiz posted Wed May 08, 2024 07:46 AM

  
Open Source AI is Broken. And at IBM, we believe we have a solution. Let me explain...

Open Source in Software is great because you tap into:

- Collaborative Improvement: a collective effort, as contributors worldwide enhance projects through shared insights and diverse expertise.

- Continuous Refinement: The code constantly evolves as community contributions are merged, leading to improved features, functionality, and performance.

Collective Ownership and a sense of pride motivate individuals to contribute towards a common goal.

BUT with AI, this isn't working.

Thousands of forks appear every time a new open-source model is released, but they rarely merge into the base model, preventing meaningful community improvements.

In simple terms, the original model is not getting more intelligent.

For example, Llama 3 was released 19 days ago, and there are already over 6,000 Llama 3 models on Hugging Face! Each represents a different version, and there's no straightforward mechanism for building upon each other.

Yesterday, we launched ๐—œ๐—ป๐˜€๐˜๐—ฟ๐˜‚๐—ฐ๐˜๐—Ÿ๐—ฎ๐—ฏ, which we believe entirely will change the game.

InstructLab addresses these challenges by enabling contributors to add specific skills or knowledge to a model.

Its model-agnostic technology allows upstream creators to regularly update their open-source models by integrating new skills instead of fully retraining them, promoting efficient and collaborative development.

InstructLab employs the LAB technique, which uses taxonomy-guided synthetic data generation and a multi-phase tuning framework to align models, making them more accessible by reducing reliance on costly human annotations.

๐Ÿคฉ We are super excited about the possibilities of InstructLab, a project that fosters collaboration by leveraging diverse expertise and perspectives to improve model alignment.

Want to learn more? Here are some links you cannot miss:

- Project: https://instructlab.ai/
- Research Paper: https://lnkd.in/e3r25GuT
- News: https://lnkd.in/eVNctKV6
- Explainer: https://lnkd.in/eJ3ZqN9A

____

graphical user interface, text, application, email


#watsonx.ai
#GenerativeAI#ai-featured-area-1

0 comments
17 views

Permalink