Group Relative Policy Optimization (GRPO)

What is Group Relative Policy Optimization (GRPO)?

Group Relative Policy Optimization (GRPO) is a reinforcement learning-based fine-tuning approach that optimizes a model using relative feedback across groups of candidate responses, rather than requiring absolute scores for individual outputs. For each input, the model generates multiple candidate responses that are evaluated by a reward function. GRPO then updates the policy by increasing the likelihood of higher-scoring responses and decreasing the likelihood of lower-scoring ones within the same group. This approach is particularly effective when you have preference data or quality comparisons between responses, and it typically produces more robust and preference-aligned model behaviors.

Prerequisites

Before installing, make sure you have:

  1. Python 3.11+

  2. Pip or

  3. gcloud CLI - required because gllm-training is a private library hosted in a private Google Cloud repository

After installing, please run

gcloud auth login

to authorize gcloud to access the Cloud Platform with Google user credentials.

Our internal gllm-training package is hosted in a secure Google Cloud Artifact Registry. You need to authenticate via gcloud CLI to access and download the package during installation.

  1. The minimum requirements:

    1. CUDA-compatible GPU

    2. Recommendation GPU:

      1. RTX A5000

      2. RTX 40/50 series.

    3. Windows/Linux, currently not support for macOS

Installation

pip install --extra-index-url "https://oauth2accesstoken:$(gcloud auth print-access-token)@glsdk.gdplabs.id/gen-ai-internal/simple/" "gllm-training"

Quickstart

Let's move on to a basic example fine-tuned using GRPOTrainer. To run GRPO fine-tuning, you need to specify a reward function, model name, column_mapping and dataset path.

# Main Code
from gllm_training import GRPOTrainer
from examples.llm_as_judge_reward_function import output_format_reward

grpo_trainer = GRPOTrainer(
    model_name="Qwen/Qwen3-0.6b",
    datasets_path="examples/grpo_csv",
    reward_functions=[output_format_reward]
)
grpo_trainer.train()

Fine tuning model using YAML file.

We can run experiments in a more structured way by using a YAML file. The current GRPO fine-tuning SDK supports both online data from Google Spreadsheets and local data in CSV format.

Example 1: Fine tuning using online data.

We can prepared our experiment using YAML file with the data trained and validation from google spreadsheet.

1

Configure environment variables (.env)

Fill in the GOOGLE_SHEETS_CLIENT_EMAIL and GOOGLE_SHEETS_PRIVATE_KEY fields. If you don’t have these keys, please contact the infrastructure team.

2

Share the spreadsheet

Share your Google Spreadsheet containing the training and validation data with the GOOGLE_SHEETS_CLIENT_EMAIL.

3

Experiment configuration (grpo_experiment_config.yml)

You can use a YAML file to plan your fine tuning experiments. To fine tuning with YAML, you need to define the required variables in the file.

4

Reward function

Reward functions evaluate model outputs and convert them into numerical feedback that GRPO uses to update the policy. In practice, a reward function:

  • Takes a batch of completions (model-generated outputs)

  • Computes a float reward for each completion.

  • Returns a list of floats where each element corresponds to exactly one completion.

5

Fine tuning

To run your fine-tuning, you need to load the YAML data using the YamlConfigLoader function, and select the experiment ID when executing the load function.

6

(Notes) Output format

Our SDK supports dictionary or string output formats for fine-tuned models.

  1. Output dictionary

    YAML format

    Expected output

  2. Output string

    YAML format

    Expected output

Example 2: Fine tuning using local data.

The remaining hyperparameter configurations for fine-tuning are the same as when using online data. Below is an example YAML configuration for using local data for training and validation.

Logging Monitoring

During the GRPO fine-tuning process, the SDK automatically generates comprehensive logs to help you monitor training progress and debug issues. These logs are stored in two formats:

JSONL Logs (Structured Training Metrics)

The SDK generates structured JSONL logs that capture detailed training metrics at each step. These logs are stored in:

Example path: data/grpo/model/exp_999/Qwen3-0.6b/logs/grpo_train_steps.jsonl

Each line in the JSONL file contains a JSON object with training metrics such as:

  • step: Training step number

  • loss: Training loss at that step

  • learning_rate: Current learning rate

  • epoch: Current epoch number

  • rewards/mean: Average reward across generated responses

  • rewards/std: Standard deviation of rewards

  • rewards/margin: Difference between best and worst rewards in the group

  • policy_loss: Policy optimization loss

  • And other relevant metrics

You can parse these logs programmatically or use tools like jq to analyze the training progression:

Note: The JSONL file contains both training steps and reward metrics. Use select() to filter for the specific type of data you need.

Tensorboard Logs (Visual Monitoring)

For visual monitoring and analysis, the SDK also generates TensorBoard-compatible logs stored in:

Example path: data/grpo/model/exp_1/Qwen3-0.6b/logs_tensorboard

To visualize your training progress:

  1. Launch TensorBoard:

  1. Open your browser and navigate to http://localhost:6006

  2. Monitor key metrics:

    • Training/Validation loss curves

    • Learning rate scheduling

    • Step-by-step progress

    • Custom metrics (if configured)

Log Configuration

You can customize logging behavior through the hyperparameters configuration:

Best Practices

  1. Monitor reward signals: Check that average rewards increase over time - if rewards stay flat or decrease, your reward function may not be aligned with your goals or the model isn't learning the desired behavior.

  2. Track policy stability: Use Tensorboard to see policy loss trends - sudden spikes in policy loss or reward variance indicate training instability and may require reducing learning rate.

  3. Debug reward function: JSONL logs show reward statistics for each step - look for reward mean trends, check if reward margin is too small (responses are too similar), or if standard deviation is too high (inconsistent quality).

  4. Monitor generation quality: GRPO generates multiple responses per input, so watch that reward mean increases while maintaining reasonable diversity (reward std shouldn't be zero).

  5. Compare experiments: Save logs from each experiment to compare which reward functions and hyperparameters produce the best policy improvements - use TensorBoard to view multiple experiments side-by-side

Upload model to cloud storage

When running experiments, we don’t always save the model directly to the cloud. Instead, we may first evaluate its performance before uploading it to cloud storage. To support this workflow, we provide a save_model function that allows you to upload the model as a separate step after fine tuning.

1

Configure environment variable (.env)

Fill in the AWS_ACCESS_KEY, AWS_SECRET_KEY and AWS_REGION fields. If you don’t have these keys, please contact the infrastructure team.

2

Upload model

To upload the model, you need to configure the storage configuration and specify the model path on save_model function. The model path should point to the directory of your best adapter model.

PreviousSupervised Fine Tuning (SFT)NextDirect Preference Optimization (DPO)

Last updated