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: |

The seemingly negative pandemic effects and resource/product shortages are encouraging and allowing organizations to innovate and change.The news of cash-rich organizations getting ready for the post-Covid growth economy is a sign of more than capital spending plans. Cash provides a cushion for risk-taking and a tool for growth.

To pay the bills, Mr. Durov is issuing investors $1 billion to $1.5 billion of company debt, with the promise of discounted equity if the company eventually goes public, the people briefed on the plans said. He has also announced plans to start selling ads in public Telegram channels as soon as later this year, as well as offering other premium services for businesses and users.

Parallel Experiments from sa


Telegram Parallel Experiments
FROM USA