added section on on-policy distillation #220
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f86b90c1b1
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| $$\mathcal{L}_{\text{KD}}(\theta) = -\,\mathbb{E}_{x \sim \mathcal{D}}\left[\sum_{t=1}^{T} P_{\phi}(x_t \mid x_{<t}) \log P_{\theta}(x_t \mid x_{<t})\right]. $$ {#eq:knowledge_distillation} | ||
| This is sometimes called **offline distillation** because the teacher generates data independently of the student. An alternative is **on-policy distillation**, where the student samples its own outputs and receives feedback from the teacher, which connects distillation to reinforcement learning through the direction of KL divergence minimization [@agarwal2024policy]. See Chapter 16 for details on on-policy distillation and its relationship to RLHF. |
There was a problem hiding this comment.
This sentence directs readers to Chapter 16 for on-policy distillation details, but Chapter 16 is the evaluation chapter and contains no on-policy distillation content, while the new section lives in Chapter 15. Readers following this cross-reference will end up in the wrong chapter and miss the material entirely. Consider updating the reference to the chapter that actually introduces on-policy distillation.
Useful? React with 👍 / 👎.
|
@kashif I appreciate the PR. I'm kind of torn about it and how I'd make edits (while being busy), so I may end up sitting on it for a bit 🙏 until i have inspiration |
|
no worries! |
2 participants
No description provided.