[Example] GRPO on GSM8K with RULER reward #239
Conversation
There was a problem hiding this comment.
Summary of Changes
Hello @hiyuchang, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request introduces a novel example demonstrating the application of Generalized Reinforcement Learning with Policy Optimization (GRPO) to solve mathematical problems on the GSM8K dataset. The core innovation lies in the adoption of RULER, an LLM-as-a-judge reward mechanism, to provide feedback for the reinforcement learning agent. This allows for the exploration of training agents using sophisticated, LLM-generated reward signals rather than traditional rule-based rewards, potentially leading to more nuanced and effective learning.
Highlights
- New Example for GRPO on GSM8K: Introduced a new example demonstrating the application of Generalized Reinforcement Learning with Policy Optimization (GRPO) to the GSM8K dataset.
- RULER Reward Integration: Implemented RULER (Relative Universal LLM-Elicited Rewards) as the primary reward function, which leverages an auxiliary LLM to act as a judge for evaluating model responses.
- MathRULERWorkflow: Added a new workflow,
MathRULERWorkflow, specifically designed to handle the RULER-based reward calculation and integrate it into the training process. This workflow also calculates a 'gold reward' for comparison. - Configuration Files: New configuration files (
gsm8k_ruler.yamlandtrain_gsm8k_ruler.yaml) have been added to define the specific parameters for this GRPO experiment, including settings for the auxiliary LLM,std_thresholdfor GRPO advantage, and learning rate. - New Metrics: Introduced new metrics for tracking, such as
reward(RULER score),gold_reward(rule-based), andjudge_success(whether RULER successfully returned a valid score).
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces a new example for GRPO on GSM8K using a RULER reward, including a new MathRULERWorkflow, configuration files, and documentation. My review has identified a critical security vulnerability in the use of eval() on LLM output. Additionally, there are several high-severity issues, such as a misconfigured file path that will prevent the example from running, a redundant initialization call, and an off-by-one error in a prompt. I have also included some medium-severity suggestions to improve code quality and clarity.
|
/unittest-module-common |
Summary
Tests
Github Test Reporter by CTRF 💚 |
Description
As the title says.
Checklist
Please check the following items before code is ready to be reviewed.