trl/pr_4624/en/quickstart.md

# Quickstart

Get started with TRL in minutes. This guide shows you the essentials for training models with SFT, GRPO, and DPO.

> 💡 **Looking for ready-to-run examples?** Check out our [notebooks for Colab](#ready-to-run-examples) or [production scripts](example_overview#scripts).

## Quick Examples

Copy-paste these minimal examples to start training immediately. Each uses compact models for quick experimentation.

### Supervised Fine-Tuning

```python
from trl import SFTTrainer
from datasets import load_dataset

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",
    train_dataset=load_dataset("trl-lib/Capybara", split="train"),
)
trainer.train()
```

### Group Relative Policy Optimization

```python
from trl import GRPOTrainer
from datasets import load_dataset
from trl.rewards import accuracy_reward

trainer = GRPOTrainer(
    model="Qwen/Qwen2.5-0.5B-Instruct",  # Start from SFT model
    train_dataset=load_dataset("trl-lib/DeepMath-103K", split="train"),
    reward_funcs=accuracy_reward,
)
trainer.train()
```

### Direct Preference Optimization

```python
from trl import DPOTrainer
from datasets import load_dataset

trainer = DPOTrainer(
    model="Qwen/Qwen2.5-0.5B-Instruct",  # Use your SFT model
    ref_model="Qwen/Qwen2.5-0.5B-Instruct",  # Original base model
    train_dataset=load_dataset("trl-lib/ultrafeedback_binarized", split="train"),
)
trainer.train()
```

### Reward Modeling

```python
from trl import RewardTrainer
from datasets import load_dataset

dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")

trainer = RewardTrainer(
    model="Qwen/Qwen2.5-0.5B-Instruct",
    train_dataset=dataset,
)
trainer.train()
```

## Command Line Interface

Skip the code entirely - train directly from your terminal:

```bash
# SFT: Fine-tune on instructions
trl sft --model_name_or_path Qwen/Qwen2.5-0.5B \
    --dataset_name trl-lib/Capybara

# DPO: Align with preferences  
trl dpo --model_name_or_path Qwen/Qwen2.5-0.5B-Instruct \
    --dataset_name trl-lib/ultrafeedback_binarized

# Reward: Train a reward model
trl reward --model_name_or_path Qwen/Qwen2.5-0.5B-Instruct \
    --dataset_name trl-lib/ultrafeedback_binarized
```

## Ready-to-Run Examples

Want to dive deeper? We provide a comprehensive collection of examples for all skill levels:

  
    Notebooks (Beginner-friendly)
    Self-contained notebooks for interactive learning. Many run on free Google Colab, while some require larger GPUs.
    
      SFT with QLoRA (free Colab ✓)
      GRPO with QLoRA (free Colab ✓)
      GRPO for Vision-Language Models (free Colab ✓)
    
    → See all notebooks
  
  
    Scripts (Production-ready)
    Full-featured scripts for single GPU, multi-GPU, and DeepSpeed setups. Ready for real-world training.
    
      SFT Script
      GRPO Script
      DPO Script
    
    → See all scripts
  

## What's Next?

### 📚 Learn More

- [SFT Trainer](sft_trainer) - Complete SFT guide
- [DPO Trainer](dpo_trainer) - Preference alignment
- [GRPO Trainer](grpo_trainer) - Group relative policy optimization

### 🚀 Scale Up

- [Distributed Training](distributing_training) - Multi-GPU setups
- [Memory Optimization](reducing_memory_usage) - Efficient training
- [PEFT Integration](peft_integration) - LoRA and QLoRA

### 🌐 Community

- [Community Tutorials](community_tutorials) - External guides and resources

## Troubleshooting

### Out of Memory?

Reduce batch size and enable optimizations:

```python
training_args = SFTConfig(
    per_device_train_batch_size=1,  # Start small
    gradient_accumulation_steps=8,  # Maintain effective batch size
)
```

```python
training_args = DPOConfig(
    per_device_train_batch_size=1,  # Start small
    gradient_accumulation_steps=8,  # Maintain effective batch size
)
```

### Loss not decreasing?

Try adjusting the learning rate:

```python
training_args = SFTConfig(learning_rate=2e-5)  # Good starting point
```

For more help, open an [issue on GitHub](https://github.com/huggingface/trl/issues).