PPO Training
Train language models using Proximal Policy Optimization (PPO) for reinforcement learning from human feedback (RLHF).
Overview
PPO training is a 2-step process:
- Train a Reward Model - Train a model to score responses (see Reward Modeling)
- Run PPO Training - Use the reward model to guide policy optimization
Quick Start
aitraining llm --train \
--model google/gemma-3-270m \
--data-path ./prompts.jsonl \
--project-name ppo-model \
--trainer ppo \
--rl-reward-model-path ./reward-model
Python API
from autotrain.trainers.clm.params import LLMTrainingParams
from autotrain.project import AutoTrainProject
params = LLMTrainingParams(
model="google/gemma-3-270m",
data_path="./prompts.jsonl",
project_name="ppo-model",
trainer="ppo",
rl_reward_model_path="./reward-model",
# PPO hyperparameters
rl_gamma=0.99,
rl_gae_lambda=0.95,
rl_kl_coef=0.1,
rl_clip_range=0.2,
rl_num_ppo_epochs=4,
epochs=1,
batch_size=4,
lr=1e-5,
)
project = AutoTrainProject(params=params, backend="local", process=True)
project.create()
Requirements
PPO training requires either --rl-reward-model-path (path to a trained reward model) or --model-ref (reference model for KL divergence). At least one must be specified.
Parameters
Core PPO Parameters
| Parameter | CLI Flag | Default | Description |
|---|
rl_reward_model_path | --rl-reward-model-path | None | Path to reward model (required) |
rl_gamma | --rl-gamma | 0.99 | Discount factor (0.9-0.99) |
rl_gae_lambda | --rl-gae-lambda | 0.95 | GAE lambda for advantage estimation (0.9-0.99) |
rl_kl_coef | --rl-kl-coef | 0.1 | KL divergence coefficient (0.01-0.5) |
rl_value_loss_coef | --rl-value-loss-coef | 1.0 | Value loss coefficient (0.5-2.0) |
rl_clip_range | --rl-clip-range | 0.2 | PPO clipping range (0.1-0.3) |
rl_value_clip_range | --rl-value-clip-range | 0.2 | Value function clipping range |
Training Parameters
| Parameter | CLI Flag | Default | Description |
|---|
rl_num_ppo_epochs | --rl-num-ppo-epochs | 4 | PPO epochs per batch |
rl_chunk_size | --rl-chunk-size | 128 | Training chunk size |
rl_mini_batch_size | --rl-mini-batch-size | 8 | Mini-batch size |
rl_optimize_device_cache | --rl-optimize-device-cache | True | Memory optimization |
Generation Parameters
| Parameter | CLI Flag | Default | Description |
|---|
rl_max_new_tokens | --rl-max-new-tokens | 128 | Max tokens to generate |
rl_top_k | --rl-top-k | 50 | Top-k sampling |
rl_top_p | --rl-top-p | 1.0 | Top-p (nucleus) sampling |
rl_temperature | --rl-temperature | 1.0 | Generation temperature |
Advanced Parameters
| Parameter | CLI Flag | Default | Description |
|---|
rl_reward_fn | --rl-reward-fn | None | Reward function: default, length_penalty, correctness, custom |
rl_multi_objective | --rl-multi-objective | False | Enable multi-objective rewards |
rl_reward_weights | --rl-reward-weights | None | JSON weights for multi-objective |
rl_env_type | --rl-env-type | None | RL environment type |
rl_env_config | --rl-env-config | None | JSON environment config |
{"text": "What is machine learning?"}
{"text": "Explain quantum computing."}
{"text": "Write a haiku about coding."}
RL Environment Types
Three environment types are available:
| Environment | Description |
|---|
text_generation | Standard text generation with reward scoring |
multi_objective | Multiple reward components combined |
preference_comparison | Compare generated responses |
Multi-Objective Rewards
Enable multiple reward signals:
params = LLMTrainingParams(
...
trainer="ppo",
rl_multi_objective=True,
rl_env_type="multi_objective",
rl_reward_weights='{"correctness": 1.0, "formatting": 0.1}',
)
Example: Full RLHF Pipeline
Step 1: Train Reward Model
aitraining llm --train \
--model google/gemma-3-270m \
--data-path ./preferences.jsonl \
--project-name reward-model \
--trainer reward \
--prompt-text-column prompt \
--text-column chosen \
--rejected-text-column rejected
Step 2: Run PPO Training
aitraining llm --train \
--model google/gemma-3-270m \
--data-path ./prompts.jsonl \
--project-name ppo-model \
--trainer ppo \
--rl-reward-model-path ./reward-model \
--rl-kl-coef 0.1 \
--rl-clip-range 0.2
Best Practices
- Start with a good base model - Fine-tune with SFT before PPO
- Use a well-trained reward model - Quality of rewards determines PPO success
- Monitor KL divergence - Too high means model is diverging too much from original
- Start with default hyperparameters - Adjust based on training dynamics
- Use small learning rates - PPO is sensitive to learning rate (1e-5 to 5e-6)
Next Steps
