fix: align GRPO prompt format with SFT training format#61
Merged
Conversation
The GRPO rollout prompt was missing the "Thought:" line and action
history that the SFT training uses. Models fine-tuned via SFT output
"Thought: ...\nAction: CLICK(...)" but the GRPO prompt didn't
prompt for this format, causing verbose free-form output that
couldn't be parsed → reward 0.0 → zero gradients.
Changes:
- Add "Thought:" and "Action:" prompt lines matching SFT format
- Add action_history parameter for step context
- Parser extracts action from "Action: ..." line before regex matching
- Parser handles JSON format {"action_type": "click", "coordinate": [x,y]}
- Debug logging of raw VLM output for zero-reward diagnosis
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes zero-reward GRPO runs caused by prompt format mismatch between SFT and GRPO. Adds Thought/Action format, JSON parsing, and debug logging.