Telegram Group & Telegram Channel
Telegram Group » Singapore » 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/sg/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: |

The lead from Wall Street offers little clarity as the major averages opened lower on Friday and then bounced back and forth across the unchanged line, finally finishing mixed and little changed.The Dow added 33.18 points or 0.10 percent to finish at 34,798.00, while the NASDAQ eased 4.54 points or 0.03 percent to close at 15,047.70 and the S&P 500 rose 6.50 points or 0.15 percent to end at 4,455.48. For the week, the Dow rose 0.6 percent, the NASDAQ added 0.1 percent and the S&P gained 0.5 percent.The lackluster performance on Wall Street came on uncertainty about the outlook for the markets following recent volatility.

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 sg


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