Add template-only overrides for RL checkpoint exports - #1905
abhishekraok wants to merge 2 commits into
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![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: fa97e8971b
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".
|
|
||
|
|
||
| ### Added | ||
| - Add an optional template-only override for MoE HF exports, a CPU-only command for existing exports, and a pinned think-dev template selected by the KDA launcher's `convert_rl` mode. Tokenizer serialization and ordinary SFT conversion defaults are unchanged. |
There was a problem hiding this comment.
When this change is submitted as a PR, the changelog CI check will reject it because the commit modifies open_instruct/ but the new entry does not contain the PR's own URL. Append the PR link to this entry once the PR exists, as required by the repository workflow.
AGENTS.md reference: AGENTS.md:L10-L12
Useful? React with 👍 / 👎.
This branch has not been deployed
1 participant
Codex · codex://threads/01a0bc67-0eac-7681-babe-a1a745942f39
The KDA SFT converter exports an instruct-dev chat template, so RL prompts do not force
<think>. Add--export-chat-templateto replace only the saved template and any embedded template metadata, plus a CPU-only command for existing HF exports. The override leaves tokenizer segmentation, special tokens, and weights untouched.The new
convert_rllauncher mode requires the saved training tokenizer directory and writes a separate-thinkexport using the checked-in think-dev template pinned tocf6d9298ed2f5d53f098f1785b58a23a24dff8ed. Ordinary conversion defaults are unchanged. This is a temporary template choice until Kevin Farhat's Olmo 3.5 template is ready; it does not change the converter's existing tokenizer serialization behavior.Stacked on kda-think-sft (#1886), which contains the converter and launcher. Addresses the RL export handoff in dolci-think-sft-anchor (#1895).
Validation: 15 CPU tests passed, covering template rendering, tokenizer/weight preservation, linked metadata, CLI integration, and launcher arguments.
make style quality, script Ruff checks, shell syntax, andgit diff --checkpassed using the existing local environment without dependency sync. Weight conversion was mocked in the integration tests; no compute was launched and no existing checkpoint was modified.GPU_TESTS=bypass — template metadata and launcher changes; no model computation changes.