Telegram Group & Telegram Channel
Telegram Group » Hong Kong » Parallel Experiments » Telegram Webview » Post 938
Parallel Experiments
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。
arXiv.org
Direct Preference Optimization: Your Language Model is Secretly a...
While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely...
www.tg-me.com/hk/Parallel Experiments/com.LinghaoCh/938
1.2K viewsLinghao Zhang, edited  



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: |

Should You Buy Bitcoin?

In general, many financial experts support their clients’ desire to buy cryptocurrency, but they don’t recommend it unless clients express interest. “The biggest concern for us is if someone wants to invest in crypto and the investment they choose doesn’t do well, and then all of a sudden they can’t send their kids to college,” says Ian Harvey, a certified financial planner (CFP) in New York City. “Then it wasn’t worth the risk.” The speculative nature of cryptocurrency leads some planners to recommend it for clients’ “side” investments. “Some call it a Vegas account,” says Scott Hammel, a CFP in Dallas. “Let’s keep this away from our real long-term perspective, make sure it doesn’t become too large a portion of your portfolio.” In a very real sense, Bitcoin is like a single stock, and advisors wouldn’t recommend putting a sizable part of your portfolio into any one company. At most, planners suggest putting no more than 1% to 10% into Bitcoin if you’re passionate about it. “If it was one stock, you would never allocate any significant portion of your portfolio to it,” Hammel says.

Launched in 2013, Telegram allows users to broadcast messages to a following via “channels”, or create public and private groups that are simple for others to access. Users can also send and receive large data files, including text and zip files, directly via the app.The platform said it has more than 500m active users, and topped 1bn downloads in August, according to data from SensorTower.Parallel Experiments from hk


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。

 Parallel Experiments TG
Webview: 938
Parallel Experiments.Telegram Webview
Parallel Experiments Telegram TG Channel
Telegram Updated:
Telegram Parallel Experiments
FROM USA