Overview
TRL builds on the Transformers ecosystem and provides trainers for methods such as supervised fine-tuning (SFT), direct preference optimization (DPO), and group relative policy optimization (GRPO). It supports distributed training integrations and parameter-efficient fine-tuning workflows.
Key features
- SFT, DPO, GRPO and related trainers
- Transformers ecosystem integration
- Distributed training support
- CLI and Python APIs
Requirements, installation and quick start
Install the Python package with `pip install trl`; follow the docs for optional Accelerate, PEFT, and hardware setup.
Usage
Choose a trainer, prepare a compatible dataset and model, configure training arguments, then run the trainer or CLI command.
Model compatibility and use cases
Designed for transformer models in the Hugging Face ecosystem; architecture and hardware support depends on the chosen method and dependencies.
License and risk notes
The repository declares Apache-2.0. Dataset, base-model, and hosted service terms are separate.
Release and maintenance
See the official repository for current releases and changes.