Published
How to best align a model for human interaction? In RLHF we first learn a proxy for the human preferences: the reward model (RM), that is later used to align a language policy. Yet the RM is an imperfect approximator of human preferences, and its prolonged usage leads inevitably to reward hacking. In this presentation we’ll review techniques developed in Gemma to mitigate this hacking and allow for longer training.
Subscribe to Google for Developers → https://goo.gle/developers
Speakers: Louis Rouillard
Products Mentioned: Gemma
Subscribe to Google for Developers → https://goo.gle/developers
Speakers: Louis Rouillard
Products Mentioned: Gemma
- Category
- Project
- Tags
- Google, developers, pr_pr: Gemma;
Sign in or sign up to post comments.
Be the first to comment
Up Next
-
01:37
How to drive clarity and alignment with Copilot in Microsoft Viva Goals
-
07:24
Demo: Gemma Scope: Sparse autoencoders on Gemma 2
-
04:12
Which Gemma version is the right one for you?
-
01:51
What is Gemma Scope?
-
06:39
Welcome to Gemma 3
-
15:50
Star Wars Jedi: Survivor - How to Solve the Alignment Control Center Screen Puzzle
-
01:13
Drive Mission and Alignment with Microsoft Viva
-
09:35
AI for everyone with Gemma
-
00:35
Top 3 use cases for retraining Gemma?
-
00:15
Run AI locally, with Gemma!
-
01:02
Hugging Face + Gemma = ????
-
00:27
What Gemma model can help me code?
-
06:31
Gemma in Minutes: 3 ways to run Gemma’s latest version!
-
12:10
Get started with Gemma 3 on Vertex AI
-
14:29
What's new in Gemma 2
-
09:02
A truly multilingual Gemma 3
-
11:29
2020 Corvette Stingray Long-Term Update ― Sports Car Performance Test! Track vs. Street Alignment
-
03:10
6 Best Derailleur Hanger Alignment Gauges 2019
-
26:48
Deepdive into Gemma 3
-
04:21
What’s new in Gemma 3?
-
05:13
Car Tech 101: Understanding wheel alignment
-
09:00
Getting started with Gemma models
Loading...