[Logging] save generations to a huggingface dataset for online trainers#364
[Logging] save generations to a huggingface dataset for online trainers#364kashif wants to merge 5 commits intothinking-machines-lab:mainfrom
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: 9869b0004a
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".
|
|
||
| if rows["step"]: | ||
| ds = Dataset.from_dict(rows) | ||
| path = self.parquet_dir / f"completions_{step:06d}.parquet" |
There was a problem hiding this comment.
In the streaming and async RL paths, log_trajectories is invoked once per trajectory group (often multiple times for the same step), but this logger always writes to completions_{step:06d}.parquet. That means later calls for the same step overwrite earlier ones instead of appending, so runs with multiple groups per step silently lose most logged generations and produce incomplete HF datasets.
Useful? React with 👍 / 👎.
1 participant
No description provided.