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: |

Mr. Durov launched Telegram in late 2013 with his brother, Nikolai, just months before he was pushed out of VK, the Russian social-media platform he founded. Mr. Durov pitched his new app—funded with the proceeds from the VK sale—less as a business than as a way for people to send messages while avoiding government surveillance and censorship.

Among the actives, Ascendas REIT sank 0.64 percent, while CapitaLand Integrated Commercial Trust plummeted 1.42 percent, City Developments plunged 1.12 percent, Dairy Farm International tumbled 0.86 percent, DBS Group skidded 0.68 percent, Genting Singapore retreated 0.67 percent, Hongkong Land climbed 1.30 percent, Mapletree Commercial Trust lost 0.47 percent, Mapletree Logistics Trust tanked 0.95 percent, Oversea-Chinese Banking Corporation dropped 0.61 percent, SATS rose 0.24 percent, SembCorp Industries shed 0.54 percent, Singapore Airlines surrendered 0.79 percent, Singapore Exchange slid 0.30 percent, Singapore Press Holdings declined 1.03 percent, Singapore Technologies Engineering dipped 0.26 percent, SingTel advanced 0.81 percent, United Overseas Bank fell 0.39 percent, Wilmar International eased 0.24 percent, Yangzijiang Shipbuilding jumped 1.42 percent and Keppel Corp, Thai Beverage, CapitaLand and Comfort DelGro were unchanged.

Parallel Experiments from us


Telegram Parallel Experiments
FROM USA