TRL v1.0: Post-Training Library Built to Move with the Field

TRL v1.0 marks a major release of the open-source post-training library, transitioning from research code to production-grade infrastructure.

• •Implements 75+ post-training methods covering PPO, DPO, ORPO, KTO, RLVR-style methods, and more, making diverse optimization approaches accessible
• •Addresses the challenge of building stable software in a constantly evolving field where core assumptions and method architectures change unpredictably
• •Separates stable and experimental APIs under one package, allowing rapid method innovation while maintaining backward compatibility guarantees
• •Deliberately limits abstractions and favors explicit implementations with acceptable code duplication over inflexible inheritance hierarchies
• •Provides broad method coverage, deep Hugging Face integration, and low infrastructure requirements while maintaining semantic versioning contracts