Telegram Group & Telegram Channel
https://arxiv.org/abs/2305.18290 #llm #ai

今天深入学习了 DPO,再次感叹扎实的数学功底对 AI/ML Research 的重要性……

原始的 RLHF 是用 pairwise human preference data(A 和 B 哪个更好)去训练一个 reward model,然后用 RL 来训练主 policy model,objective 是 minimize negative log likelihood + regularization(比如 PPO 就是通过新旧 policy 之间的 KL Divergence 来做 regularization)。这样的缺点在于 RL 是出了名的难搞,而且还需要一个 critic model 来预测 reward,使得整个系统的复杂性很高。

DPO 的思路是,观察到 RLHF 的 objective 本质上是 minimize loss over (latent) reward function,通过一番 reparameterization 等数学推导,重新设计了一个 minimize loss over policy 的 objective,绕过了中间这个 reward model,让 gradient update 直接增加 policy model 生成 winner response 的概率并降低 loser response 的概率,大幅简化了流程。

拓展阅读:
- KTO: 更进一步,不需要 pairwise comparison,只用对 individual example 的 upvote/downvote 也可以学习到 preference。
- IPO: 解决 DPO 容易 overfit 的问题。



tg-me.com/LinghaoCh/938
Create:
Last Update:

https://arxiv.org/abs/2305.18290 #llm #ai

今天深入学习了 DPO,再次感叹扎实的数学功底对 AI/ML Research 的重要性……

原始的 RLHF 是用 pairwise human preference data(A 和 B 哪个更好)去训练一个 reward model,然后用 RL 来训练主 policy model,objective 是 minimize negative log likelihood + regularization(比如 PPO 就是通过新旧 policy 之间的 KL Divergence 来做 regularization)。这样的缺点在于 RL 是出了名的难搞,而且还需要一个 critic model 来预测 reward,使得整个系统的复杂性很高。

DPO 的思路是,观察到 RLHF 的 objective 本质上是 minimize loss over (latent) reward function,通过一番 reparameterization 等数学推导,重新设计了一个 minimize loss over policy 的 objective,绕过了中间这个 reward model,让 gradient update 直接增加 policy model 生成 winner response 的概率并降低 loser response 的概率,大幅简化了流程。

拓展阅读:
- KTO: 更进一步,不需要 pairwise comparison,只用对 individual example 的 upvote/downvote 也可以学习到 preference。
- IPO: 解决 DPO 容易 overfit 的问题。

BY Parallel Experiments




Share with your friend now:
tg-me.com/LinghaoCh/938

View MORE
Open in Telegram


Parallel Experiments Telegram | DID YOU KNOW?

Date: |

Telegram announces Search Filters

With the help of the Search Filters option, users can now filter search results by type. They can do that by using the new tabs: Media, Links, Files and others. Searches can be done based on the particular time period like by typing in the date or even “Yesterday”. If users type in the name of a person, group, channel or bot, an extra filter will be applied to the searches.

Export WhatsApp stickers to Telegram on Android

From the Files app, scroll down to Internal storage, and tap on WhatsApp. Once you’re there, go to Media and then WhatsApp Stickers. Don’t be surprised if you find a large number of files in that folder—it holds your personal collection of stickers and every one you’ve ever received. Even the bad ones.Tap the three dots in the top right corner of your screen to Select all. If you want to trim the fat and grab only the best of the best, this is the perfect time to do so: choose the ones you want to export by long-pressing one file to activate selection mode, and then tapping on the rest. Once you’re done, hit the Share button (that “less than”-like symbol at the top of your screen). If you have a big collection—more than 500 stickers, for example—it’s possible that nothing will happen when you tap the Share button. Be patient—your phone’s just struggling with a heavy load.On the menu that pops from the bottom of the screen, choose Telegram, and then select the chat named Saved messages. This is a chat only you can see, and it will serve as your sticker bank. Unlike WhatsApp, Telegram doesn’t store your favorite stickers in a quick-access reservoir right beside the typing field, but you’ll be able to snatch them out of your Saved messages chat and forward them to any of your Telegram contacts. This also means you won’t have a quick way to save incoming stickers like you did on WhatsApp, so you’ll have to forward them from one chat to the other.

Parallel Experiments from us


Telegram Parallel Experiments
FROM USA