fix(typecheck): stabilize GRPO setup and scored-data typing#42
fix(typecheck): stabilize GRPO setup and scored-data typing#42Dexploarer wants to merge 5 commits intoJejuNetwork:developfrom
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: c36be009b7
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".
| const nonce = bytesToHex(randomBytes(32)) | ||
| const timestamp = Date.now() |
There was a problem hiding this comment.
The external flow generates a fresh nonce/timestamp (and therefore a different message) every time jeju login --external runs, but the UX is explicitly two-step (print message first, submit signature later). On the second invocation, verifyMessage is run against a new message, so signatures from the printed message cannot validate and external login cannot complete.
Useful? React with 👍 / 👎.
| `jeju login --network ${network} --address ${address} --signature <your-signature>`, | ||
| ) |
There was a problem hiding this comment.
The follow-up command shown to users omits --external, but signature handling only runs inside the if (options.external) branch. If users follow this instruction literally, the CLI skips external-signature verification and falls back to private-key auth, so the documented external flow fails.
Useful? React with 👍 / 👎.
1 participant
Summary\n- Fix optional model attribute access in GRPO trainer setup by using local typed model refs before methods/properties are called.\n- Refactor Jeju RLAIF env scored-group typing to avoid invalid TypedDict inheritance and keep the inference logprobs payload typed via a local alias/cast.\n\n## Validation\n- bunx pyright packages/training/python apps/dws/api