Demo: Gemma Scope: Sparse autoencoders on Gemma 2

Channel:
United States Google for Developers
Subscribers:
2,510,000
Published on ● Video Link: https://www.youtube.com/watch?v=X1gDXDQu_wU


Duration: 0:00
302 views
9

Sparse Autoencoders (SAEs) are an interpretability tool that allow you to inspect what concepts your LLM is thinking about. Learn more our open-source suite of SAEs on all layers of Gemma 2 2B and 9B, which enables everyone to analyze and steer Gemma's internals.

Watch more videos from Gemma Developer Day Tokyo → https://goo.gle/gemma-dev-tokyo

Subscribe to Google for Developers → https://goo.gle/developers

#Gemma #GemmaDeveloperDay



Other Videos By Google for Developers


2024-10-18Demo: Post-training research with Gemma
2024-10-18Demo: Gemma 2 architecture: JAX, Flax, and more
2024-10-18Fine-tuning Gemma for the world's languages
2024-10-18Demo: Gemma on-device with MediaPipe
2024-10-18Fireside chat on Gemma's expansion to non-English languages
2024-10-18What's new in Gemma 2
2024-10-18Demo: Gemma data, training, and the path to improvement
2024-10-18Demo: Multi-modal Gemma: PaliGemma with Keras
2024-10-18Demo: Teach Gemma to work in YOUR language
2024-10-18Demo: DataGemma: Grounding LLMs with Data Commons data
2024-10-18Demo: Gemma Scope: Sparse autoencoders on Gemma 2
2024-10-17Demo: Unleashing Gemma in production with Hugging Face Text Generation Inference (TGI)
2024-10-17Demo: Gemma 2 2B on a Jetson Orin Nano
2024-10-17How to run and scale Gemma 2 on Google Cloud
2024-10-16What Gemma model can help me code?
2024-10-15Personal AI Code Assistant with Gemma | Build with Google AI
2024-10-15AI Code Assistant with no third-party services required
2024-10-13Revolutionize how your team interacts with data!
2024-10-09Run AI locally, with Gemma!
2024-10-09Think you can hack Google? Lets find out.
2024-10-08What is Google’s Vulnerability Rewards Program?