Repository navigation
feat: Phase 4.B on-policy vision RL - #57
Merged
Merged
Conversation
- run_grpo accepts ModelInput prompts and per-prompt rewards; passes detokenized text into reward_fn; job=vision_grpo metrics + n_image_refs - vision_rewards: keyword, phrase, action-bin, weighted rubric scorers - run_vision_grpo + run_vision_rl_queue curriculum on shared LoRA - catalog vision_rl_on_policy; smoke scripts/vision_rl_demo.py
3 tasks done
ajwharton
added a commit
that referenced
this pull request
Jul 30, 2026
Keep Phase 4.B as done on main; retain FPS smoke + dogfood next steps.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Phase 4.B product path for on-policy vision/robot RL under four verbs:
run_grpoprompts may beModelInput(text tokens +ImageRefChunk)vision_rewards)run_vision_rl_queue(caption → hazard ladder toy)Also fixes reward_fn always receiving
""(now gets detokenized completion when provided).Test plan
python scripts/vision_rl_demo.py --steps 2