Reinforcement Learning: ChatGPT and RLHF
HTML-код
- Опубликовано: 16 июн 2024
- Reinforcement Learning from human feedback, and how it's used to help train large language models like ChatGPT.
Part 3 of RL from scratch series.
• Reinforcement Learning...
0:00 - intro
0:06 - large language models
0:35 - learning to tell jokes
1:13 - fine tuning with better data
1:26 - positive and negative examples
2:03 - reinforcement learning for LLMs
3:00 - labeling fewer examples
3:56 - reward networks
5:08 - summing it up
5:23 - variants
5:57 - chatGPT, Bard, Claude, Llama
6:09 - finally, a good joke!
Please come back, your videos are great!
All of your videos are amazing, please upload more
Joel, excellent explanation and talk! Thank you!
Welcome back!
Hope to see more of these videos..
Super helpful - thank you for this series!
Amazing content! Please keep them coming!
help me a lot, can't wait to see more
Great content!!
Good teaching.
Who is this guy? He made all the complexity so simple with his words. Anyone know this gentleman name?
You are the Best
How long it takes to train a reward network? And how reliable would it be?
come back :(
🎯 Key Takeaways for quick navigation:
00:00 🤖 Reinforcement learning improves large language models like ChatGPT.
00:25 🃏 Large language models face issues like bias, errors, and quality.
01:11 📊 Training data quality impacts results; removing bad jokes might help.
01:55 🧩 Training on both good and bad jokes improves language models.
02:38 🔄 Language models are policies, reinforcement learning uses policy gradient.
03:08 🎯 Reinforcement Learning from Human Feedback (RLHF) challenges data acquisition.
03:35 🤔 RLHF theory: Language model might already know jokes' boundary.
04:18 🏆 Training a reward network predicts human ratings for model's output.
04:47 🔄 Reward network is a modified language model for predicting ratings.
05:14 📝 Approach: Humans write text, train reward network, refine model with RL.
05:57 ⚖️ Systems convert comparisons to ratings for reward network training.
06:11 😄 RLHF successfully improves language models, including humor.
Made with HARPA AI