Skip to content

Update reward terminology from 'shaped' to 'dense' #175

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
qgallouedec wants to merge 1 commit into
base: feature/gpro-rlvr-recipes
Choose a base branch
from
Open

Update reward terminology from 'shaped' to 'dense' #175

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions 0_model_customization_recipes/preference_optimization/grpo_rlvr/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,7 +169,7 @@ def accuracy_reward(completions: List[List[Dict]], answer: List[str], **kwargs)
### Reward Design Tips

- **Sparse rewards** (0.0 or 1.0): Simple but can be slow to learn
- **Shaped rewards** (0.0 to 1.0): Provide intermediate feedback
- **Dense rewards** (0.0 to 1.0): Provide intermediate feedback
- Partial credit for correct tool selection
- Partial credit for correct argument types
- Full credit for correct final answer
Expand Down Expand Up @@ -366,7 +366,7 @@ The `GRPOTrainer` (in `grpo_trainer_v2.py`):
## Tips

1. **Start simple**: Begin with 2-3 tools and exact-match rewards
2. **Iterate on rewards**: Experiment with shaped rewards for faster learning
2. **Iterate on rewards**: Experiment with dense rewards for faster learning
3. **Validate tools**: Test your tool functions independently before training
4. **Monitor rewards**: Watch mean reward per batch to track learning
5. **Use clear docstrings**: The model sees your function docstrings as tool descriptions
Expand All @@ -383,7 +383,7 @@ The `GRPOTrainer` (in `grpo_trainer_v2.py`):

**Low rewards throughout training**
- Check that expected answers match tool output format exactly
- Try shaped rewards with partial credit
- Try dense rewards with partial credit
- Verify tools are being called (check logs)

**Model not calling tools**
Expand Down